Learning To Use Statistical Tests in Psychology

Learning to use
statistical tests
in psychology
L
e
a
r
n
i
n
g

t
o

u
s
e

s
t
a
t
i
s
t
i
c
a
l
t
e
s
t
s

i
n

p
s
y
c
h
o
l
o
g
y
T H I R D E D I T I O N
J U D I T H G R E E N E
M A N U E L A D O L I V E I R A
G
R
E
E
N
E

A
N
D

D
O
L
I
V
E
I
R
A
LEARNING TO USE STATISTICAL TESTS IN PSYCHOLOGY
THIRD EDITION
Praise for the rst edition:
An excellent textbook which is well planned, well written, and pitched at the
correct level for psychology students. I would not hesitate to recommend Greene
and dOliveira to all psychology students looking for an introductory text on
statistical methodology.
Bulletin of the British Psychological Society
Learning to Use Statistical Tests in Psychology third edition has been updated
throughout. It continues to be a key text in helping students to understand and
conduct statistical tests in psychology without panic! It takes students from the
most basic elements of statistics teaching them:
How psychologists plan experiments and statistical tests
Which considerations must be made when planning experiments
How to analyze and comprehend test results
Like the previous editions, this book provides students with a step-by-step guide
to the simplest non-parametric tests through to more complex analysis of
variance designs. There are clear summaries in progress boxes and questions for
the students to answer in order to be sure that they have understood what they
have read.
The new edition is divided into four sections and within this structure each test
covered is illustrated through a chapter of its own. The sections cover:
The principles of psychological research and psychological statistics
Statistical tests for experiments with two or three conditions
Statistical tests based on ANOVA (Analysis of Variance) conditions as well
as tests for multiple comparisons between individual conditions
Statistical tests to analyze relationships between variables
Presented in a student-friendly textbook format, Learning to Use Psychological
Tests in Psychology enables readers to select and use the most appropriate
statistical tests to evaluate the signicance of data obtained from psychological
experiments.
Judith Greene was educated at Oxford University and University
College London where she obtained her PhD in psychology. After ten years
lecturing at Birkbeck College London, she was appointed Professor of Psychology
at the Open University in 1976, where she is now Emeritus Professor.
Manuela dOliveira did her rst degree at Birkbeck College London
and obtained her PhD at the Open University, where she worked from 1976 to
1985. She is now Invited Professor at the Faculty of Psychology and Educational
Sciences, University of Porto, Portugal.
3
r
d
E
d
i
t
i
o
n
Learning to use statistical
tests in psychology
THIRD EDITION
Learning to use statistical
tests in psychology
THIRD EDITION
Judith Greene
and
Manuela DOliveira
Open University Press
Open University Press
McGraw-Hill Education
McGraw-Hill House
Shoppenhangers Road
Maidenhead
Berkshire
England
SL6 2QL
email: enquiries@openup.co.uk
world wide web: www.openup.co.uk
and Two Penn Plaza, New York, NY 10121-2289, USA
First published 1982
Reprinted 1983, 1985, 1987, 1988, 1989, 1990, 1992, 1993, 1994, 1995, 1996, 1998
First published in this second edition, 1999
Reprinted 2000, 2001
First published in this third edition, 2005
Copyright Judith Greene and Manuela DOliveira, 2005
All rights reserved. Except for the quotation of short passages for the purpose
of criticism and review, no part of this publication may be reproduced, stored
in a retrieval system, or transmitted, in any form or by any means, electronic,
mechanical, photocopying, recording or otherwise, without the prior written
permission of the publisher or a licence from the Copyright Licensing Agency
Limited. Details of such licences (for reprographic reproduction) may be
obtained from the Copyright Licensing Agency Ltd of 90 Tottenham Court
Road, London, W1T 4LP.
A catalogue record of this book is available from the British Library
ISBN-13: 978 0335 21680 3 (pb) 978 0335 21681 1 (hb)
ISBN-10: 0335 21680 3(pb) 0335 21681 1 (hb)
Library of Congress Cataloging-in-Publication Data
CIP data has been applied for
Typeset by ReneCatch Limited, Bungay, Suffolk
Printed in the UK by ???????????????????
To Norman, Kate and Matthew
and
To Pedro, Maria Joo, Catarina, Miguel, Mariana, Carolina
Contents
Preface to the third edition xii
Study guide for students xiv
Acknowledgements xvi
Part I Introduction 1
1 Research in psychology 3
1.1 Psychological research and statistics 3
1.2 Variability in human behaviour 3
1.3 Relationships between variables 4
1.4 Research hypothesis 4
1.5 Null hypothesis 4
1.6 Rationale for using statistical tests 5
1.7 Participants in psychological research 5
2 Experiments in psychology 7
2.1 The experimental method 7
2.2 Independent and dependent variables 7
2.3 Experimental and control conditions 8
2.4 Three or more experimental conditions 9
2.5 Same participants (related designs) 10
2.6 Different participants (unrelated designs) 10
3 Selecting statistical tests 12
3.1 Basic principles for selecting tests 12
3.2 Experiments 12
3.3 Number of experimental conditions 13
3.4 Related or unrelated designs 13
3.5 Non-parametric or parametric tests 13
3.6 Using the Decision Chart 13
Contents viii
4 Using statistical tests 15
4.1 Variability in data 15
4.2 Probabilities in statistical tables 17
4.3 Selecting a level of signicance 17
4.4 One-tailed and two-tailed hypotheses 19
Part II Statistical tests for experiments 21
5 Introduction to non-parametric tests 23
5.1 Ordinal data 23
5.2 Assigning ranks to scores 23
5.3 Assigning tied ranks 24
5.4 Standard headings for presenting statistical tests 25
6 Wilcoxon test (related) 27
6.1 Scores from same participants 27
6.2 Ranking differences between conditions 28
6.3 Ranking plus and minus differences 29
6.4 Ties between scores 29
6.5 Selecting a statistical test in the Decision Chart 30
6.6 Using the Wilcoxon test (related) 31
7 MannWhitney test (unrelated) 36
7.1 Scores from different participants 36
7.2 Overall ranking of scores 37
7.3 Rank totals for each condition 37
7.5 Using the MannWhitney test (unrelated) 39
8 Introduction to parametric t tests 44
8.1 Comparisons between parametric and non-parametric tests 44
8.2 Numerical calculations for parametric tests 45
8.3 Calculating variances 46
8.4 Ratio of variances 46
9 t test (related) 48
9.2 Squaring differences 48
9.5 Using the t test (related) 50
10 t test (unrelated) 55
10.2 Squaring scores 55
10.5 Using the t test (related) 57
Contents ix
11 Friedman test (related) 62
11.1 Scores from same participants for three conditions 62
11.2 Ranking three conditions 63
11.3 Assigning tied ranks 64
11.4 Rank totals 65
11.6 Using the Friedman test (related) 66
12 KruskalWallis test (unrelated) 72
12.1 Scores from different participants for three (or more)
conditions 72
12.2 Overall ranking of scores 73
12.3 Totals of ranks for each condition 74
12.5 Using the KruskalWallis test (unrelated) 75
12.6 Note about names of non-parametric tests 79
Part III Analysis of variance 81
13 Introduction to ANOVA 83
13.1 Parametric tests 83
13.2 Analysis of variance 83
13.3 Sources of variance 84
13.4 Degrees of freedom 85
13.5 Requirements for parametric tests 86
13.6 Normal distribution 87
13.7 Computer statistical packages 88
14 One-way ANOVA (unrelated) 90
14.2 Denitions of variance 90
14.4 Using one-way ANOVA (unrelated) 91
15 One-way ANOVA (related) 96
15.1 Scores from same participant 96
15.2 Denitions of variance 96
15.4 Using one-way ANOVA (related) 97
16 Comparisons between ANOVA conditions 103
16.1 Overall differences in ANOVA 103
16.2 Graphs showing patterns of mean scores 103
16.3 Comparisons between individual conditions 104
16.4 Multiple comparisons between conditions 105
16.5 Using t tests for pairs of conditions 105
16.6 Adjusting signicance levels 106
Contents x
16.7 Bonferroni test 107
16.8 Tukeys honestly signicant difference 107
17 Introduction to two-way ANOVA 111
17.1 Comparison with one-way ANOVA 111
17.2 Two-way ANOVA 111
17.3 Two-by-two tables 112
17.4 Interpreting interactions 114
17.5 Using graphs to demonstrate interactions 115
18 Two-way ANOVA (unrelated) 118
18.2 Denitions of variances 119
18.4 Using two-way ANOVA (unrelated) 120
18.5 Using graphs 125
19 Two-way ANOVA (related) 127
19.2 Denitions of variances 128
19.4 Using two-way ANOVA (related) 129
19.5 Using the SPSS computer program 137
19.6 Note about mixed designs 138
Part IV Relationships between variables 141
20 Chi-square 143
20.1 Comparison with experiments 143
20.2 Measuring categories 143
20.3 Nominal data 144
20.4 Predicting categories 144
20.6 Using the chi-square test 146
21 Pearson product moment correlation 152
21.1 Comparison with experiments 152
21.2 Measuring variables in correlations 152
21.3 Predicting positive correlations 153
21.4 Predicting negative correlations 154
21.5 Correlation coefcients 154
21.7 Using Pearson correlation product moment correlation 155
22 Introduction to simple linear regression 161
22.1 Comparisons with correlation 161
22.2 Predictor variable and criterion variable 162
22.3 Scatterplots 163
Contents xi
22.4 Regression equation for calculating a regression line 164
22.5 Residual error 166
22.6 Rationale for using ANOVA as a test of linear regression 166
22.7 ANOVA table to test linear regression 167
23 Multiple regression 169
23.1 Comparisons with simple regression 169
23.2 Objectives of multiple regression 169
23.3 Types of multiple regression 172
23.5 Using multiple regression 174
24 General linear model 183
24.1 Introduction 183
24.2 The role of error in GLM 183
24.3 Application of the GLM to linear regression 184
24.4 Application of GLM to ANOVA 184
24.5 Conclusions 184
Answers to questions 186
Recommended reading 195
References 196
Statistical tables 197
Index 209
Decision Chart: inside back cover
Preface to the third edition
There have been an enormous number of textbooks claiming to present
statistics in a simple way. Despite this, many psychology students still nd
the whole statistical business something of a mystery. How does this book
differ from these other attempts?
We believe that virtually all books on statistics feel obliged to start with
the mathematical principles underlying probability distributions, samples
and populations, and statistical testing. But, however simply these are
presented, in our view they obscure the basic reasons why psychologists
use statistical tests.
So we had better come clean straight away. This book sets out to achieve
one aim. This is to enable students to select appropriate statistical tests
to evaluate the signicance of data obtained from psychological research.
In other words, this book is concerned with inferential statistics as used in
psychological studies.
We have concentrated on this to the exclusion of much else. Topics such
as descriptive statistics and the basic principles of probability are well
covered in other statistical texts. There is nothing here about the use of
surveys, observational techniques or psychometric tests of intelligence and
personality. All we have included is the battery of statistical tests which are
usually introduced to psychology students as part of their undergraduate
laboratory course. We hope that, by aiming at a single target, we will
maximize our chances of scoring a bulls-eye.
This beginners book takes students from the simplest non-parametric
tests, like the Wilcoxon test, through to complex analysis of variance
designs. The principle is the same throughout: always to give the rationale
for using appropriate statistical analyses for particular research designs.
It is certainly our expectation that anyone who has mastered the why
and how of the statistical tests given in this book will be in a position to
understand the basic principles of statistical tests as presented in advanced
textbooks of psychological statistics.
Preface to the third edition xiii
Our belief is that, with the aid of this book, students will feel comfort-
able about the basis for selecting and applying all kinds of statistical tests.
We hope that teachers will wish to use a book which frees students from
much of the panic usually associated with statistics. That way they should
be in a far more receptive state to learn.
One major change in this third edition is that the book is now organized
into four parts. Part I consists of a general introduction to the principles
of psychological research and psychological statistics. Part II describes
statistical tests for straightforward experiments with two and three con-
ditions. Part III introduces the statistical tests based on ANOVA designs.
It also considers tests for multiple comparisons between individual
conditions.
Part IV is a new part. By omitting some of the less used non-parametric
tests, we have been able to bring together the statistical tests suitable for
analysing relationships between variables, including correlation and
regression.
Another innovation is to devote a whole chapter to each statistical test,
making it easier to draw attention to the special requirements for each
test. These additions will make the book more comprehensive in line with
the range of tests generally available.
Study guide for students
The aim of this book is to explain the rationale for using statistical tests
to evaluate the results of psychological research. The problem is that in
psychology you have to carry out these experiments on human beings,
often other students. Unlike most physical objects, human beings are
unique, each interpreting and performing whatever task you set them in a
slightly different way.
You will nd that the data and observations obtained from the people
doing a psychological experiment are often extremely varied and that
many of the things which inuence their behaviour may have nothing to
do with the aims of the research. It is for this reason that you have to sort
out whether experimental results are really signicant. And, as you will
see, this is just what statistical tests enable you to do.
You will probably be relieved to hear that the chapters which introduce
the basic rationale for statistics and summarize all you need to know in
order to select an appropriate statistical test are the shortest chapters in the
book. These are aspects of using statistical tests that students often nd
rather puzzling, but we hope that these chapters will clear up all your
worries.
Other chapters in the book present statistical tests, explaining the
rationale for each one, taking you step by step through any necessary
calculations and giving precise instructions about how to use the statistical
tables at the back of the book. For more complex types of statistical
analysis you will be introduced to the latest types of computer programs
for carrying out numerical calculations.
One essential feature of this book is the questions which occur through-
out the text. It is not enough to read the summaries presented in the
progress boxes. The only way you can make sure that you understand
the context of each section is to attempt the questions before looking up
the answers at the back! It is equally important to work your way through
the step-by-step instructions given for each statistical test. Otherwise, you
Study guide for students xv
will never gain the condence which comes from fully understanding the
rationale for a statistical test.
Let us end by making some important and, we hope, encouraging
points. The rst thing to grasp is that statistical tests are not magic for-
mulae to be turned to, desperately wondering which on earth to choose.
They simply follow as a natural result of the kind of research you have
chosen to do. What makes most people give up all hope of mastering
statistics is the thought that they will nd themselves presented with a huge
amount of numerical data without the foggiest idea of how to deal with
them. But this is quite the wrong way to go about things. The important
thing is to decide what research you want to carry out. You will nd that
such a decision immediately narrows your possible choice of statistical
tests to only one or two, and that there are good reasons for selecting one
or the other.
With statistical tests selection is all; the actual calculations are quite easy
once you have understood the reasons for doing them. The aim has been
to introduce the principles of using statistical tests without referring to
any mathematical concepts. And, in order to do the calculations for the
tests themselves, you will only need to know how to add, subtract, multi-
ply, divide and square numbers. With modern pocket calculators and
computer programs this really should be childs play.
Good luck and if, in spite of everything, you do nd yourself getting
disheartened by statistics, turn back and reread this study guide.
Acknowledgements
Grateful acknowledgement is made to the following sources for per-
mission to reprint the tables in this book:
Table A from F. Wilcoxon and R.A. Wilcox, Some Rapid Approximate
Statistical Procedures, American Cyanamid Co., 1949;
Table B from R.P. Runyon and A. Haber, Fundamentals of Behavioral
Statistics, 3rd edn, Reading, Mass., Addison-Wesley, 1976;
Table C from D.V. Lindley and J.C.P. Miller, Cambridge Elementary
Statistical Tables, 10th edn, Cambridge University Press, 1973;
Table D from M. Friedman, The use of ranks to avoid the assumption of
normality implicit in the analysis of variance in Journal of the American
Statistical Association, Vol. 32, 1937;
Table E from D.V. Lindley and J.C.P. Miller, Cambridge Elementary
Statistical Tables, 10th edn, Cambridge University Press, 1973;
Table F from W.H. Kruskal and W.A. Wallis, Use of ranks in one-
criterion variance analysis in Journal of the American Statistical
Association, Vol. 47, 1952;
Table G from D.V. Lindley and J.C.P. Miller, Cambridge Elementary Stat-
istical Tables, 10th edn, Cambridge University Press, 1973
Table H from Table VII (p. 63) of R.A. Fisher and F. Yates, Statistical
Tables for Biological, Agricultural and Medical Research (6th edition),
published by Longman Group Ltd, London, 1974 (previously published
by Oliver and Boyd Ltd, Edinburgh).
Introduction Part I
Research in psychology 1
Psychological research and statistics
1.1
You may be wondering why a book on learning to use statistical tests in
psychology would start with an introduction to the general principles of
psychological research. You may be pleased to hear that psychologists do
not study statistics for its own sake. Statistical tests simply provide a tool
for analysing the results of psychological research.
The aim of psychological research is to test psychological theories about
human behaviour. Researchers make predictions about human behaviour
in order to test whether the results of the research support the prediction
or not.
Once you appreciate that the point of psychological research is to test
research predictions, you will nd that learning to select and use statistical
tests will be reasonably straightforward.
Variability in human behaviour
1.2
One of the particular aspects of humans is that they vary a great deal in
their characteristics and behaviour. So it is difcult for psychologists to
make predictions that will apply to all human beings in exactly the same
way. Even identical twins develop differently, depending on circumstances.
For this reason, psychological researchers make predictions in a way
that reects the variability in human behaviour.
The variability between humans is formulated as variables. The aim of
psychological research is to study variables in human behaviour.
Introduction 4
Relationships between variables
1.3
The major implication for psychological research is that predictions are
made about relationships between variables in human behaviour.
Let us take an example of measuring childrens scores for reading and
arithmetic. One thing is certain. All the children will not have exactly the
same reading scores. Prociency in reading will be very variable. For this
reason, reading scores represent a variable.
The same would apply to the childrens arithmetic scores. Again the
arithmetic scores are likely to vary from child to child over quite a wide
range. For this reason, arithmetic scores represent a variable.
The researcher would be interested in testing whether the children who
have high scores on reading also have high scores on arithmetic. The aim
would be to discover whether there is a relationship between these two
variables. One of the two variables is the variable of reading scores and the
second variable is the variable of arithmetic scores.
Research hypothesis
1.4
In psychological research a prediction about relationships between
variables is formulated as a research hypothesis. A research hypothesis
might predict that there will be a relationship between reading scores and
arithmetic scores. The aim of psychological research is to test whether the
research hypothesis is supported by the scores of the children on both
variables.
The research hypothesis predicts a relationship between variables. It is
called a hypothesis because it is something that is proposed and has to be
tested.
Null hypothesis
1.5
One especially important point is that, in order to test a research hypothesis,
it must be possible for the predicted relationships between variables to
be supported or not to be supported. If it were always certain that the
prediction would be supported there would be no point in doing psycho-
logical research.
In our example, the predicted relationship between the variables of
reading scores and arithmetic scores may or may not be supported by the
childrens scores on these two variables. It may turn out that the childrens
scores do not support the predicted relationship between reading and
arithmetic.
Research in psychology 5
This is a basic rule of psychological research. A research hypothesis
is always tested against a null hypothesis, which states that a researcher
will not nd the research results he/she expects. According to the null
hypothesis, any results found in a research study are due to random vari-
ability in the childrens scores rather than the predictions that the
researcher is interested in testing.
Rationale for using statistical tests
1.6
The next question is how to set about showing whether there is indeed a
predicted relationship between the variables of childrens reading scores
and arithmetic scores. The issue is whether it is possible to rule out the null
hypothesis.
Because human behaviour is variable, it is not easy to rule out the
possibility that the reading and arithmetic scores are random scores.
Perhaps the scores of the children taking part in the research are so
variable that the null hypothesis of random variability in scores cannot be
rejected.
The rationale for using statistical tests is to measure how much vari-
ability there is in childrens scores. Do the scores support the research
hypothesis that reading scores and arithmetic scores are related? Or is
there just too much random variability in the scores?
Psychologists use statistical tests to decide between these two outcomes.
Do the scores support the research hypothesis or are the scores random, as
stated by the null hypothesis?
Participants in psychological research
1.7
This is a good point at which to introduce a new terminology to describe
the people taking part in psychological research. These people are called
participants. This term is use to indicate that the people are actively
participating in psychological research.
An older term used for the people taking part in psychological research
(especially in experiments) was subjects. This was originally used in order
to differentiate human subjects from the objects used in scientic research.
Although the term subjects has now been replaced by participants,
you will still come across subjects in older articles in psychological
journals and in some textbooks on statistics. We will be using the term
participants and will only refer to subjects as an alternative term when
the term subject is always used.
Introduction 6
Progress box one

Psychologists study human beings, who are variable in their
behaviour.
Psychologists make predictions about relationships between variables,
formulated as a research hypothesis.
This is contrasted with a null hypothesis, which states the possibility
that the predicted relationships will not occur because of random
variability in scores.
Statistical tests are used to decide between the research hypothesis
and the null hypothesis.
The people taking part in psychological research are called
participants.
Experiments in psychology 2
The experimental method
2.1
In experimental research, as in all psychological research, experimenters
make predictions about relationships between variables when predicting
human behaviour. They have to take into account the variability in
humans so that their predictions are probabilities about how participants
will react rather than certainties.
The special feature of experiments is that an experimenter selects one of
the variables to test whether this variable has any effect on a second variable.
An example would be an experimenter giving children a reading scheme
and measuring the effects of the reading scheme on the childrens scores
on a reading test. These are both variables. One variable varies between
giving a reading scheme and not giving a reading scheme. The childrens
scores are a variable because the scores will vary from child to child.
The research hypothesis predicts the effects of the reading scheme
variable on the variable of childrens scores. The research hypothesis has to
be compared with the null hypothesis that there are random differences in
the childrens scores.
Independent and dependent variables
2.2
We will now introduce the terminology of independent variables and
dependent variables in experimental research. The variable that the
researcher decides to test is known as the independent variable. This is
because the reading scheme and no reading scheme conditions are
selected independently by the experimenter before the experiment even
begins.
The second variable, representing childrens scores, is known as the
dependent variable. This is because the reading scores are dependent on the
Introduction 8
way in which the experimenter rst selects the independent variable of the
reading scheme.
In order to test the predictions in the research hypothesis the experi-
menter decides to vary an independent variable to see whether this has any
effect on the dependent variable. In the reading scheme experiment the
researcher selects one variable (reading scheme or no reading scheme) to see
what effect this will have on the second variable (childrens reading scores).
The aim is to test the predicted independent variable and to attempt to rule
out other irrelevant variables that might also be affecting participants
behaviour. The remainder of this chapter will discuss various ways in
which experimenters attempt to eliminate the effects of irrelevant variables.
Experimental and control conditions
2.3
The basic experimental design is for the researcher to allocate participants
to experimental conditions representing the independent variable. A com-
parison is made between participants scores on the dependent variable.
For instance, children would be allocated to the reading scheme condition
or the no reading scheme condition and their reading scores would be
measured.
The problem is that all the children might show an improvement in
reading scores over time, regardless of whether they had been given a
reading scheme or not. Can the experimenter be certain that any improve-
ment in reading is due to the reading scheme?
Experimenters always try to rule out irrelevant variables that are not
predicted by the research hypothesis. In order to achieve this, experi-
menters try to design an experiment in which the only difference between
the two conditions is due to the predicted effects of the independent
variable formulated in the research hypothesis.
It is for this reason that experimenters often introduce a control
condition against which the effects of an independent variable can be
compared. The point about a control condition is that it represents a
condition in which people are not subjected to the experimental condition.
So it is possible to compare two conditions, one with the reading scheme
and one without the reading scheme. The experimental design would be:
Experimental condition
Reading scheme Reading scores
Control condition
No reading scheme Reading scores
The prediction would be that there is more improvement in the reading
scores for the experimental condition than for the control condition. The
researcher might nd some general improvement in scores for both con-
ditions due to other variables such as a new teacher. However, if there is a
more improvement in the experimental condition, it can be claimed that
this extra improvement is due, not to factors that might have affected both
conditions equally, but to the reading scheme which affected only the
experimental condition and not the control condition.
Sometimes, rather than comparing an experimental and a control
condition, it is more appropriate to compare two levels of an independent
variable. An example would be an experiment to investigate whether lists
of common words are easier to remember than lists of rare words. The
experimental design would compare recall scores under two experimental
conditions:
Experimental condition 1
Learning lists of common words Recall of words
Learning lists of rare words Recall of words
?
Suppose a researcher is testing a research hypothesis that common words Question 2.1
are more easily recalled than rare words.
(a) What is the independent variable?
(b) What is the dependent variable?
(c) Would higher recall scores in Experimental condition 1 support the
research hypothesis?
Three or more experimental conditions
2.4
So far we have been talking about comparing only two experimental
conditions. But researchers might want to look at more than two con-
ditions to test an independent variable. For instance, they might be inter-
ested in investigating whether learning a list of vary rare words, a list of
fairly common words and a list of very common words would have an
effect on participants ability to recall the lists of words:
Learning lists of very rare words
Learning lists of fairly common words
Introduction 10
Learning lists of very common words
You would still be testing one independent variable, but with three experi-
mental conditions. The research hypothesis would be that participants
would have better recall scores for the list of very common words.
Same participants (related designs)
2.5
The next important decision to be made is whether to use the same parti-
cipants for all conditions or whether to use groups of different participants
in each condition. As we shall see, both these options have advantages and
disadvantages.
From the point of view of reducing individual differences between
participants, there is a case for using the same participants for all
conditions. The rationale is that, when using the same participants,
everything about the participants will be the same in all conditions. If
one individual participant is highly motivated when learning one list
of words, he/she will also be highly motivated when learning other lists of
words.
Using the same participants is known as a related design. Because the
comparison is within participants, the scores from each participant are
related.
The main disadvantage of using the same participants for all conditions
is that all participants are exposed to all the experimental conditions. The
experimenter has to take into account the possibility that participants may
be affected by their experience of earlier conditions.
In the attempt to reduce order effects, it is normal practice for
researchers to vary the order of conditions so that half the subjects learn
list A rst and list B second and the other half learn list B rst and list A
second. The idea is that any irrelevant effects of learning from experience
are likely to be equalized by reversing the order of the conditions. This is
known as counterbalancing the order of conditions.
Different participants (unrelated designs)
2.6
Another option is to use groups of different participants for each experi-
mental condition. There are some cases when it is obvious that different
participants are required. Imagine that you are interested in whether boys
or girls are more likely to benet from a reading scheme. There is simply
no way in which the same children could be allocated to the boy or girls
group. There would have to be different groups of girls and boys.
The main advantage of using different participants for each condition is
that each participant is only exposed to one condition. Because of this
there can be no order effects of one condition on another condition. Using
different participants is always the best option if it is considered that
exposure to one condition may have an effect on another condition.
Using different participants is known as an unrelated design. Because
the comparison is between participants the scores from each participant
are unrelated.
The problem is how to allocate different participants to each of the
experimental conditions. Perhaps all the participants who are the best
learners might be allocated to one condition. Alternatively, all the partici-
pants who cannot be bothered may be allocated to another group.
One method is to try and match the participants. The experimenter
might give participants a prior memory test in order to identify the
best learners. The experimenter could then make sure to allocate equal
numbers of good learners and not so good learners to each group.
Another method is to allocate participants at random to each of
the conditions. If participants are randomly allocated to experimental
conditions, then people of different ages or abilities are just as likely to be
found in all the experimental groups.
All this goes to show how many things have to be taken into account
when designing an experiment. The point is to try to eliminate all possible
irrelevant variables that might be affecting participants. The only dif-
ferences between the two conditions should be the predicted differences
due to the independent variable.
Progress box two

Experimenters decide on an independent variable and predict the
effect of this variable on the dependent variable of participants
scores.
Experimenters select experimental conditions to test the research
hypothesis.
Experimenters attempt to rule out irrelevant variables that are not
predicted by the research hypothesis.
Advantages of related deigns are fewer individual differences
between the same participants doing all experimental conditions.
Experimenters need to counterbalance possible order effects.
Unrelated designs are suitable when it is important that different
participants should only do one experimental condition. Individual
differences between participants can be offset by matching partici-
pants or by randomly allocating participants to different conditions.
Selecting statistical tests 3
Basic principles for selecting tests
3.1
Now we come to the all important chapter as far as your ability to select
statistical tests is concerned. You should be encouraged to note that this is
one of the shortest chapters in the book. The vital consideration is to
choose an appropriate statistical test for each type of research. The whole
art of using statistics is to match up research designs with the statistical
tests listed in the remaining chapters of this book
We have constantly emphasized that there is a lot of variability in
participants scores caused by individual differences. This means that
every research hypothesis has to be compared against a null hypothesis
that there is random variability in scores. Each statistical test will be appro-
priate for designs that produce different amounts of random variability.
So the rst and most important point we want to emphasize is this: the
selection of an appropriate statistical test follows from the design you have
chosen to test your research hypothesis. Choice of a suitable test depends
on a few decisions about research designs. Once you have made these
decisions you will nd that you have automatically selected the appropriate
statistical test to use for analysing your research data.
Experiments
3.2
Many of the statistical tests described in this book are designed to test the
results of experiments. Experiments test predictions about differences
between conditions as a result of the experimenter deciding to test an
independent variable.
In later chapters you will be introduced to other statistical tests, which
are suitable for testing other types of relationships between variables.
Selecting statistical tests 13
Number of experimental conditions
3.3
An important aspect of a research design is how many experimental
conditions there are in an experiment.
Some statistical tests are designed to analyse experiments with only two
conditions.
Experimental designs with three or more conditions require more
complex statistical comparisons between conditions.
Related or unrelated designs
3.4
One of the most important criteria for selecting a statistical test depends
on whether the same participants have been used for all conditions (related
design) or whether groups of different participants have been used for each
condition (unrelated design).
Non-parametric or parametric tests
3.5
There are two main kids of statistical tests described in this book, called
non-parametric tests and parametric tests. The way these tests are identied
depends on how the scores (known as data) are measured.
Non-parametric tests are based on ordinal data. Ordinal data refers
to scores that can be ranked in order from lowest score to highest
score.
Parametric tests are based on interval data. Interval data refers to
scores in which the intervals between scores are equal, making it possible
to carry out numerical calculations.
You do not need to worry about this brief introduction to these different
types of tests. Details will be given for every type of research described in
this book.
Using the Decision Chart
3.6
The Decision Chart (inside back cover) consists of a set of questions you
will have to answer in order identify an appropriate statistical test.
For each statistical test questions will be asked, with suggested answers.
In every case, if you answer these questions in the Decision Chart you will
inevitably discover the correct statistical test.
Introduction 14
Progress box three

There are a few easy decisions to be made in order to select a
statistical test.
All the appropriate statistical tests are contained in the Decision
Chart.
For each type of research appropriate questions are asked in the
Decision Chart.
Using statistical tests 4
Variability in data
4.1
However much experimenters try to eliminate irrelevant variables in their
experimental designs, they can never get rid of all the variability in par-
ticipants scores. This variability in scores will inevitably be reected in
whatever measure you take of the dependent variable. What you want to
know is this: are the differences in the scores the result of manipulating
the independent variable? Or is there only random variability in the scores,
as stated by the null hypothesis?
Suppose the research hypothesis is that participants will have better
memory recall for simple texts than for complex texts. The independent
variable consists of giving one group of participants a simple text to read
and another group a complex text. The number of ideas recalled is the
measure of the dependent variable.
After running the experiment you may end up with recall scores for
each condition, as shown in Table 4.1. Total scores are given for each
condition. The means represent average scores calculated by dividing
the totals by the number of participants, in this case 10 participants
in each condition. It is clear that the data in Table 4.1 show that mean
scores for number of ideas recalled is higher in Condition 1 than in
Condition 2.
?
Does the difference between the means support the research hypothesis? Question 4.1
There would be no problem about claiming support for the hypothesis if
each and every participant in Condition 1 had recalled exactly 6.4 ideas
while every participant in Condition 2 had recalled 3.6 ideas. In that case it
Introduction 16
would be obvious that the participants in Condition 1 recalled more ideas
than in Condition 2.
But life is not that easy. There may be a lot of variability in scores
that has nothing to do with the predicted differences between simple and
complex texts.
If you look at the scores in Table 4.1 you will notice that participants
scores varied in the two conditions. For instance, although the scores were
generally higher in Condition 1, there was one participant in Condition 1
who recalled only 3 ideas in the simple text. Many participants scored
more ideas in the complex text.
?
(a) In Table 4.1 how many participants in both conditions recalled 5 ideas? Question 4.2
(b) What were the lowest and highest scores in Condition 1?
(c) What were the lowest and highest scores in Condition 2?
It is obvious from the scores in Table 4.1 that there is quite a lot of
variability in the number of correct ideas recalled. Clearly, means do not
tell us everything we need to know about participants scores. We also need
a measure of the total variability of scores.
We now get to the heart of the rationale for statistical tests, which is to
measure the amount of variability in the data. Does the difference between
the means indicate signicant support for the research hypothesis?
Table 4.1 Number of ideas recalled
Condition 1
(simple texts)
Condition 2
(complex texts)
10 2
5 1
6 3
3 4
9 4
8 4
7 2
5 5
6 7
5 4
Totals 64 36
Means 6.4 3.6
Alternatively, are the differences between conditions swamped by the total
variability in scores, as stated by the null hypothesis?
Probabilities in statistical tables
4.2
A big difference in mean scores between conditions may be due to the
predicted effects of the independent variable rather than random vari-
ability. But there is always a specic probability that the differences in
scores are caused by total random variability. So there can never be 100 per
cent certainty that the scores in an experiment are due to the effects of
selecting the independent variable.
Statistical tests calculate probabilities that results are signicant. The
way this is done may seem rather paradoxical. Statistical tables provide
probabilities that any differences in scores are due to random variability, as
stated by the null hypothesis. This means that the less probable it is that
any differences are due to random variability, the more justication there is
for rejecting the null hypothesis.
This is the basis of all statistical tests. Statistical tables give the
probability that scores in an experiment occur on a random basis. If
the probability that the scores are random is very low, then you can reject
the null hypothesis that the differences are random. Instead you can
accept the research hypothesis that the experimental results are signicant,
that is, that they are not likely to be random.
Strictly speaking, the only conclusion from looking up probabilities
in statistical tables is that they justify rejecting the null hypothesis. But
you will nd that, if the null hypothesis can be rejected, psychological
researchers usually claim that the results provide support for the predic-
tions in the research hypothesis.
Selecting a level of signicance
4.3
So far we have not really tackled the question of why a researcher would
be prepared to accept a particular percentage probability of random
variability when deciding whether or not to reject the null hypothesis.
What risk would be acceptable that the scores resulting from the experi-
ment occurred randomly, as stated by the null hypothesis, and were not
signicant at all? Of course, you would like to be 100 per cent certain
that the difference in scores is signicant. But, as we said earlier, you
can never be 100 per cent certain that the total scores are not randomly
distributed.
Would you accept a 99 per cent probability that your results are signi-
cant against a 1 per cent probability that they are random? Or would you
Introduction 18
accept a 95 per cent probability of signicance against a 5 per cent prob-
ability that the scores are random?
These percentage probabilities are known as levels of signicance. There
is no simple answer as to what level of signicance is acceptable. It is up
to experimenters to decide what odds they are prepared to accept when
deciding whether the results of an experiment are signicant. For this
reason, experimenters always have to state the percentage probability of
random variability that they will accept for rejecting the null hypothesis.
Imagine that you are investigating whether a new reading scheme might
help children to read. You carried out an experiment in which you com-
pared the progress of a group of children using your new scheme against a
control group using traditional methods. Suppose the probability that this
difference could have occurred randomly was 5 per cent. This represents
1 in 20 chance that the scores are random. Would you accept that the
difference was signicant and introduce the new reading scheme, and at
what cost in materials and teacher training?
Imagine another case when you are testing a powerful drug with nasty
side-effects and nd an improvement in patients taking the drug as
compared with a control group. If the difference between the groups could
have occurred by chance 5 in 100 times, would you accept that the dif-
ference is signicant and introduce the new drug? Would you change your
odds if you knew that without the drug most of the patients would die
anyway? And how would you feel if an aeroplane you were going to y in
had a 5 per cent probability of developing electrical failure?
These examples bring home the fact that choosing a signicance level is
always a matter of deciding what odds you are prepared to accept that
your results are due to random variability. In psychology (possibly because
it is thought that nothing too terrible can happen as a result of accepting a
result as signicant!) there is a convention to accept probabilities of either
1 per cent or 5 per cent as grounds for rejecting the null hypothesis.
The way levels of signicance are expressed is to state that the prob-
ability of a result being due to random variability is less than 1 per cent
or less than 5 per cent. That is why in articles in psychological journals
you will see statements that differences between experimental conditions
are signicant (p < 0.01) or signicant (p < 0.05). This means that the
probability (p) of a result occurring by chance is less than (expressed as <)
1 per cent (0.01) or 5 per cent (0.05).
Sometimes, you will nd other probabilities quoted, such as p < 0.02 or
p < 0.001. These represent probabilities of obtaining a random result 2
times in 100 and 1 time in 1000 (2 per cent and 0.1 per cent). These
percentage probabilities give you grounds for rejecting the null hypothesis
that your results are due to the effects of random variability.
?
Which of the following percentage probabilities represents the smallest Question 4.3
probability of random scores?
1 in 100 (p < 0.01), 1 in 20 (p < 0.05), 1 in 1000 (p < 0.001)
We have so far been discussing cases when the probability of random
variability is quite low. But sometimes there may be a probability of
random variability that is too high to justify rejecting the null hypothesis.
A probability of 10% is considered to be too high a probability of random
variability to be able to reject the null hypothesis. In this case, the mull
hypothesis cannot be rejected so there is no signicant support for the
research hypothesis.
One-tailed and two-tailed hypotheses
4.4
There is one further point about the way an experimental hypothesis
is formulated which has implications for the way in which you look up
probabilities in statistical tables. This is whether the research hypothesis is
a one-tailed or a two-tailed hypothesis.
A one-tailed hypothesis is one that, as its name implies, makes a pre-
diction in one particular direction. An example might be that more ideas
will be recalled in a simple text than in a complex text.
With a one-tailed hypothesis the results will only be signicant if they
are in the predicted direction. If a signicant difference is found in the
opposite direction that more ideas are recalled from complex texts this
will not count, because it was not predicted.
A two-tailed hypothesis makes a prediction that the effect of an
independent variable may go in either direction. The prediction would be
that more ideas might be recalled either from a simple text or from a
complex text, that is, predicting that simple or complex texts will have
some effect on recall but not being prepared to say what.
A two-tailed hypothesis is much vaguer than a one-tailed hypothesis.
There are also differences in the probabilities of obtaining these results.
The probability of a one-tailed hypothesis is less than the probability of a
prediction in either direction for a two-tailed hypothesis.
You will nd in statistical tables that there are different probabilities for
one-tailed and two-tailed hypotheses. Because a two-tailed hypothesis
makes two predictions, one in each direction, the levels of signicance have
to be stricter. A one-tailed hypothesis has a specic percentage probability,
but for a two-tailed hypothesis there is double the probability in either
direction.
Introduction 20
Do not worry about this distinction between one-tailed and two-tailed
hypotheses at this stage. You will be given detailed instructions about what
to do for each statistical test.
Progress box four

Experimenters predict differences in scores between conditions.
There will always be random variability in scores that is not due to
the predicted differences.
Statistical tables provide probabilities of random variability, as stated
by the null hypothesis.
If the probability of random variability is low the null hypothesis can
be rejected,
Experimenters select a level of signicance, usually p < 0.05 or
p < 0.01.
If the data attains these levels of signicance the null hypothesis is
rejected and instead the data are claimed as supporting the research
hypothesis.
A one-tailed hypothesis predicts results in one direction only. A
two-tailed hypothesis predicts results in either direction.
Statistical tests for experiments Part II
Introduction to non-parametric tests 5
Ordinal data
5.1
The statistical tests described in this chapter are non-parametric tests. For
each type of statistical test there is a type of measurement that is suitable
for that test. In the case of non-parametric statistical tests it is ordinal data
that are suitable. The main feature of ordinal data is that the scores can be
ranked in order from the lowest score to the highest score.
Non-parametric tests comparing participants scores in experimental
conditions are made on the basis of whether scores can be ranked as
higher or lower in different conditions.
Ranking the scores in two experimental conditions from lowest score to
highest score is a good way of testing a research hypothesis about whether
the scores in one of the experimental conditions are ranked higher than
the scores in the other condition.
In the rest of this chapter we will be giving some general tips for ranking
scores. The methods of ranking for particular non-parametric tests will be
introduced when those tests are described.
Assigning ranks to scores
5.2
In order to establish which scores are higher or lower it is necessary to rank
scores in terms of their relative size from lowest to highest. In order to
rank scores you have to assign ranks of 1, 2, 3, 4, etc. to each score in order
of their magnitude, starting with the smallest score. The most usual way of
ranking scores is to give the lowest rank to the lowest score and the highest
rank to the highest score. This makes it easier to see which are the lowest
scores and which are the highest scores.
You will notice that we have assigned ranks to all the scores in Table 5.1.
The rst step is to identify the lowest score in Table 5.1, which is 3. We
Statistical tests for experiments 24
allocate a rank of 1 to the lowest score of 3. The next lowest score of 4 is
allocated a rank of 2. The next lowest score of 5 is ranked as 3. The rest of
the ranks are allocated in the same way, ending up with the highest score
of 12 being assigned the highest rank of 7. A good tip for calculating
number or ranks is to check that there are seven scores in Table 5.1. This
means that there should be seven ranks in all.
?
In Table 5.1 what rank has been assigned to the score of 7, and why? Question 5.1
Assigning tied ranks
5.3
We now need to consider any problems that might rise from assigning tied
ranks to identical scores. In Table 5.1 each score was different from all
the other scores. This made it easy to allocate a single rank to each of the
scores.
But in some sets of scores some of the participants may have exactly
the same recall scores. For example, one participant may have a score
of 4 and another participant also has the same score of 4. These two
scores of 4 are identical. How is it possible to assign tied ranks to these
two scores?
There are special rules for dealing with cases when there are identical
scores. If you look at the scores in Table 5.2, you will see that three par-
ticipants have recall scores of 1 and two participants have both scored 4.
What ranks should be assigned to the three scores of 1? Should these
scores all be given rank 1 because they are the lowest scores?
The basic procedure is to assign the three scores the average of the ranks
they would have been entitled to. The three scores of 1 in Table 5.2 would
have been assigned ranks 1, 2 and 3 because these are the three lowest
scores. All three scores are given the average of these ranks: 1 + 2 + 3 = 6,
Table 5.1 Ranking scores
Scores Ranks
6 4
3 1
12 7
4 2
7 5
5 3
8 6
Introduction to non-parametric tests 25
and 6 3 = 2. This result is that the three scores of 1 are each assigned the
average rank of 2.
Because the ranks of 1, 2 and 3 have already been used up for the three
lowest scores of 1, the next available rank of 4 is allocated to the next
lowest score of 2. The next rank of 5 is allocated to the next lowest score
of 3.
But what is to be done about the two identical scores of 4? The two
scores of 4 would have been entitled to the next available ranks of 6 and 7.
Since the average of 6 and 7 is 6.5 the two scores of 4 are both assigned
ranks of 6.5.
The next score of 5 will be assigned the next available rank of 8 and the
highest score of 6 gets the highest rank of 9.
Especially when there are complicated tied ranks, it is a useful tip
to count the total number of scores (in this case, 9 scores). The aim is to
check whether the highest rank of 9 has been correctly assigned to the
highest score of 6.
Standard headings for presenting statistical tests
5.4
When we describe the non-parametric statistical tests for two conditions
in Chapters 6 and 7, we shall be presenting each test under a standard set
of headings. In fact, we will be presenting all the statistical tests in this
book under the same set of headings. The aim is to make it simple for
you to follow the rationale and detailed instructions required for each
statistical test and how to look up the probabilities in statistical tables.
The standard headings for each statistical test are given below:
1 When to use. In this section you are reminded of the experimental
designs for which you can use the test.
Table 5.2 Assigning tied ranks
Scores Ranks
1 2
2 4
1 2
4 6.5
1 2
3 5
4 6.5
6 9
5 8
2 Research hypothesis. This states the prediction that the experiment is
designed to test.
3 Sample data. This gives a table of sample data that can be analysed
using the test.
4 Rationale. Under this heading the aims of the test are explained and
information is provided to help you understand the reasons for
accepting or rejecting the null hypothesis.
5 Step-by-step instructions. This section goes through a worked example
using the sample data with a clear statement of all the steps and
calculations you need to carry out when doing the test.
6 Looking up probabilities in statistical tables. This section helps you to
nd your way around statistical tables in order to decide whether the
data justify rejecting the null hypothesis.
7 Conclusions. This section states the results of the experiment and
whether they are signicant at a specic level of signicance.
Progress box ve
Non-parametric statistical tests are based on ordinal data.
Ordinal data are ranked from the lowest score to the highest score.
There are special rules for assigning tied ranks to identical scores.
Statistical tests are presented under standard headings.
Wilcoxon test (related) 6
Scores from same participants
6.1
In related designs the same participants do both conditions. Each par-
ticipant produces a pair of scores, one score for each condition. This
makes it possible to make a direct comparison between each participants
scores in the two conditions. Because each participant has a pair of related
scores in both conditions, the differences between the pairs of scores can
be calculated for each participant.
For the non-parametric Wilcoxon test the recall scores are measured
as ordinal data, which can be ranked. Each participant will have a pair of
scores, and we can calculate the differences between these pairs of ordinal
scores. Once these differences have been calculated, instructions about how
these differences can be ranked are given in Section 6.2.
For presenting sets of paired related scores it is a convention to list
the participants in the left-hand column. This is to indicate that each
participant has a pair of related scores for each condition, as shown in
Table 6.1.
We have calculated the differences for participants 1 and 2 in
Table 6.1.
?
(a) Calculate the differences for participants 3 and 4 and add these to Question 6.1
Table 6.1.
(b) Calculate the mean scores for the two conditions and add the means
to Table 6.1.
Ranking differences between conditions
6.2
As we pointed out in Chapter 5, non-parametric tests are based on ordinal
data. The differences that were calculated in Table 6.1 were between
pairs of scores, measured as ordinal data. This means that the differences
themselves are ordinal data and have to be ranked.
If you refer back to Chapter 5, Section 5.3 gives instructions about how
to assign tied ranks to identical scores. These rules apply just as much to
tied ranks in differences between scores. Ranks of the differences between
the scores in two conditions are shown in Table 6.2.
In order to rank the differences in the Differences column in Table 6.2
you need to give the lowest rank of 1 to the lowest difference of 2 and the
rank of 2 to the next lowest difference of 3.
Note the two highest differences, which are both 4. The rule is to
assign the average of the ranks these two differences would have been
entitled to. These would have been the two highest ranks of 3 and 4. The
average is 3.5 so the two differences of 4 are both allocated tied ranks
of 3.5.
Table 6.1 Differences between pairs of related scores
Participants Condition 1 Condition 2
Differences
between scores
1 10 6 4
2 5 2 3
3 4 2
4 5 2
Total scores 24 12
Mean scores
Table 6.2 Ranking differences between scores in two conditions
Participants Condition 1 Condition 2 Differences Ranks
1 6 2 4 3.5
2 5 3 2 1
3 7 4 3 2
4 7 3 4 3.5
Ranking plus and minus differences
6.3
The next point to consider is the possibility that not all the differences will
go in the same positive direction. In all the tables presented so far the
differences have always been in the same direction.
Participants may produce variable scores. It is by no means certain
that all participants will produce higher scores in Condition 1 than in
Condition 2. It is quite likely that one or other of the participants may
produce scores that are higher in Condition 2.
It is a convention that the Condition 2 scores are subtracted from the
Condition 1 scores. When there are higher scores in Condition 1 the dif-
ferences will be allocated plus signs. When the scores in Condition 1 are
lower than in Condition 2 the differences will be allocated minus signs.
The calculation of plus and minus signed differences is shown in
Table 6.3. Participants 1 and 2 had higher scores in Condition 1, but
participants 3 and 4 had lower scores in Condition 1 and higher scores in
Condition 2. By subtracting Condition 2 scores from Condition 1 scores,
participants 1 and 2 have plus differences and the participants 3 and 4 have
minus differences.
There is a special rule for ranking plus and minus differences. The rule is
that the differences are ranked in order from lowest to highest, completely
ignoring the plus and minus differences. In Table 6.3 the lowest difference
of +1 is assigned a rank of 1. The next lowest difference of 3 is assigned a
rank of 2.
Note that the two highest differences are +5 and 5. The two highest
ranks of 3 and 4 are averaged to give tied ranks of 3.5 to both these
differences.
Ties between scores
6.4
We have so far discussed applying the usual rules for assigning tied ranks
to the plus and minus differences between pairs of scores. In every case it is
Table 6.3 Ranking plus and minus differences
Differences
(1 2)
Ranks of
differences
1 10 5 +5 3.5
2 8 7 +1 1
3 7 10 3 2
4 3 8 5 3.5
the differences between pairs of scores in the two conditions that are being
ranked.
But this assumes that it is possible to calculate differences between pairs
of scores for every participant. There is a different type of tie between
pairs of scores that makes it impossible to calculate a difference between
the scores in the rst place.
This type of tie occurs when the two scores for a particular participant
are the same. There is no difference between the pairs of scores in favour
of one condition or the other. In Table 6.4 participant 3 has the same score
of 7 in both conditions.
With this type of tie between a participants paired scores on the two
conditions, the problem is that there is no basis for deciding whether this
participant had a higher score for Condition 1 or Condition 2. The only
difference that can be calculated for participant 3 is zero, indicating that
these is no difference between the two scores of 7 (shown by the 0 in the
Differences column). This lack of difference has to be recorded as a tie. As
you can see in Table 6.4, this tie must be ignored as far as the ranking of
differences is concerned.
Only four ranks can be assigned to the other four participants, rank 1
for the lowest score of 2, tied ranks of 2.5 for the two scores of +4 and 4,
and rank 4 for the highest score of 5.
Selecting a statistical test in the Decision Chart
6.5
We will now take you through the answers you need to give in the Decision
Chart (inside back cover) in order to select an appropriate statistical
test for analysing ordinal data for two conditions in which the same
participants do both conditions.
Differences between conditions?
The answer to this question is Yes because the aim of the
experiment is to test predicted differences between experimental
conditions.
Table 6.4 Ties between scores
Participants Condition 1 Condition 2 Differences (1 2) Ranks of differences
1 10 15 5 4
2 8 10 2 1
3 7 7 0 (tie)
4 6 2 +4 2.5
5 3 7 4 2.5
One or two variables?
Follow the line for one variable.
How many experimental conditions?
The answer is two conditions because the experiment is testing
differences between two conditions.
Related or unrelated design?
The answer is related design because the same participants are doing
both conditions.
Non-parametric or parametric test?
The answer is non-parametric because non-parametric statistical
tests are suitable for analysing ranked ordinal data.
If you proceed down the Decision Chart in this way, you will arrive at
the Wilcoxon test (related). The Wilcoxon test (related) is the appropriate
non-parametric test for a related design with two conditions.
Section 6.6 will provide practical advice about what you have learned
so far in this chapter about how to rank positive and negative dif-
ferences and how to deal with ties between differences and ties between
scores.
Using the Wilcoxon test (related)
6.6
When to use
The Wilcoxon test (related) should be used for a related design when the
same participants are doing two conditions and the data are ordinal.
Research hypothesis
The experimenter predicted that students would be more relaxed when
listening to music and so would recall more words in a music condition
than in a condition with no music.
?
Is this a one-tailed or a two-tailed hypothesis? Question 6.2
Sample data
All participants learned one list with music and one list without music.
To control for order effects, half the participants learned rst with music
and then without music, while the other half of the particpants did the
conditions in the reverse order. Recall scores are shown in Table 6.5.
Notes on Table 6.5
1 In the d column plus and minus differences have been calculated by
subtracting scores in Condition 2 from scores in Condition 1.
2 In the Ranks of d column ranks have been assigned regardless of the
plus and minus differences (note that there are a lot of tied ranks).
3 Participant 8 had scores of 4 for both conditions. This counts as a tie
between scores and is left out of the ranking of differences.
4 In the Signed ranks (plus) and Signed ranks (minus) columns the rank
totals for the original plus and minus differences are added up separately
to give +4 and 41, respectively.
5 The smaller rank total of (+)4 represents the value of W.
Rationale
The aim of the Wilcoxon test (related) is to compare the number of
plus and minus signed ranks in the two conditions. It is possible to
calculate the differences in the scores for each participant. If there are
only random differences between the conditions, as stated by the null
Table 6.5 Recall scores with and without music
Participants
Condition 1
(without
music)
Condition 2
(music) d Ranks
of d
Signed
ranks
(plus)
Signed
ranks
(minus)
1 3 5 2 5 ()5
2 4 5 1 2 ()2
3 3 2 +1 +2 (+)2
4 1 5 4 8.5 ()8.5
5 5 4 +1 +2 (+)2
6 2 5 3 7 ()7
7 3 5 2 5 ()5
8 4 4 0 (tie)
9 1 5 4 8.5 ()8.5
10 3 5 2 5 ()5
Means 2.9 4.5 Rank totals (+)4 ()41
hypothesis, the plus and minus differences will be roughly equal. If there is
a preponderance of plus or minus ranks in the predicted direction the null
hypothesis can be rejected.
Step-by-step instructions for calculating the value of W
These are given in Box A.
Looking up the signicance of W in Table A
Table A (at the back of the book) gives the probabilities for testing one-
tailed and two-tailed hypotheses. The smaller rank total of W = 4 has to be
equal to or less than the values in Table A. N refers to the number of
participants. In our example there were ten participants but only nine
showed a difference between the conditions. Participant 8 was a tie and so
is omitted from the analysis because the tie cannot be ranked as a plus or a
minus. So N = 9.
In Table A look along the row for N = 9. The calculated W of 4 is
smaller than the value of 6 in the second column. If you look at the top of
Box A
Step-by-step instructions for calculating the value of W
Calculate the difference (d) between 1
each pair of scores, assigning plus or
minus signs.
Subtract condition 2 from condition 1:
see the d column in Table 6.5.
Rank the differences (d) from the 2
smallest (rank 1) to the largest.
See the Ranks of d column in
Table 6.5.
When ranking omit tied scores and
ignore the plus and minus signs.
Assign plus and minus signs to the 3
ranks in the Signed ranks columns.
See plus and minus signs in Signed
ranks (plus) and Signed ranks
(minus) columns in Table 6.5.
Add the rank totals for signed ranks 4
(plus) and signed ranks (minus)
separately.
Rank total (plus) = (+)4
Rank total (minus) = ()41
Take the smaller rank total as W. 5 W = 4
Note the number of participants N 6
(not counting ties).
N = 10 1 (tie) = 9
the table the level of signicance for a one-tailed test is (p < 0.025). Since
this is a lower probability than the acceptable (p < 0.05) the null hypothesis
can be rejected.
Conclusions
It is essential to look back to the means in Table 6.5 to check whether the
difference is in the direction predicted by the one-tailed hypothesis. The
means show that participants recall more words when music is playing in
Condition 2 than without music in Condition 1 (p < 0.025).
Progress box six

In related designs pairs of scores in the two conditions are produced
by the same participants.
Differences are calculated between pairs of scores for each
participant.
These differences are ranked following the rules for assigning tied
ranks to identical differences.
Difference rated as plus or minus are ranked regardless of plus or
minus signs.
Ties between a participants scores in the two conditions, which do
not indicate a difference in favour of either condition, are recorded as
ties and cannot be ranked.
Totals of plus and minus ranks are added separately. The lower rank
is W.
The appropriate non-parametric statistical test for analysing
related ordinal data for two conditions is the Wilcoxon test
(related).
?
A research hypothesis predicted that comprehension scores would be Question 6.3
higher for a text with short sentences than long sentences. The results are
shown in Table 6.6.
(a) Calculate the mean scores for the two conditions to two decimal
places.
(b) Following the step-by-step instructions in Box A, complete the blank
columns for d, Ranks of d and the Signed ranks (plus and minus) in
Table 6.6.
(c) Is the smaller W signicant and, if so, at what level of signicance?
Table 6.6 Data for Wilcoxon (related)
Participants
Condition 1
(short
sentences)
Condition 2
(long
sentences)
d Ranks
of d
Signed
ranks
(plus)
Signed
ranks
(minus)
1 7 8
2 10 6
3 13 4
4 8 4
5 7 7
6 8 10
7 6 3
8 10 3
Means Rank totals
MannWhitney test (unrelated) 7
Scores from different participants
7.1
In unrelated designs different participants do each condition. Because
different participants do only one condition, it is not possible to compare
the scores on both conditions for each participant. Instead all individual
scores have to be ranked overall across both conditions.
In Table 7.1 scores are given for each of the conditions. But in this case
there is no column for participants. The reason is that the scores in the two
conditions are produced by two groups of different participants. We have
indicated this by referring to the two different groups who are doing the
two conditions.
For the non-parametric MannWhitney test the scores are measured as
ordinal data, which can be ranked. Instructions about how to rank all the
individual scores are given in Section 7.2.
Table 7.1 Scores for two conditions produced by different participants
Condition 1 Condition 2
Group 1 Group 2
10 5
15 6
25 7
8 10
Means 14.5 7
?
Is it possible to calculate the differences between the scores in Condition 1 Question 7.1
and Condition 2?
Overall ranking of scores
7.2
There is no basis for comparing differences in scores when groups of
different participants are doing the two conditions. But it is still necessary
to rank the scores in the two conditions in order to see whether there are
higher ranks in one condition and lower ranks in the other condition.
The way this is done is to rank all the scores in the two conditions
together. The lowest score across the two conditions is assigned a rank of
1, the next lowest gets rank 2, and so on until the highest rank is assigned
to the highest score. This results in a overall ranking of all the scores
regardless of which scores are in which condition.
The overall ranks for the scores in Table 7.1 are shown in Table 7.2. The
lowest score of 5, in Condition 2, is assigned a rank of 1. The next lowest
score of 6, also in Condition 2, is assigned a rank of 2. The third lowest
score of 7, again in Condition 2, is assigned the next rank of 3. The fourth
lowest score of 8, in Condition 1 this time, is assigned a rank of 4.
You will notice that in the overall scores there are two scores of 10. One
score of 10 was produced by a participant in Condition 1 and the other
score of 10 by a participant in Condition 2. These two scores are assigned
tied ranks of 5.5 in the overall ranks.
Rank totals for each condition
7.3
In order to decide whether the ranks are higher for one of the conditions,
the totals of ranks are added separately for each condition. The aim is to
Table 7.2 Overall ranking of scores
Scores Ranks Scores Ranks
10 5.5 5 1
15 7 6 2
25 8 7 3
8 4 10 5.5
check whether the ranks of scores in Condition 1 are higher than the ranks
of scores in Condition 2.
To make this absolutely clear, there are two stages in assigning ranks to
scores from different participants. The rst stage is to carry out an overall
ranking of all individual scores across both conditions. The second stage is
to add the rank totals for each of the conditions separately.
The totals of the ranks for Condition 1 and Condition 2 are shown in
Table 7.3.
In Table 7.1 the means are higher for Condition 1 than for Condition 2.
In Table 7.2 the higher scores in Condition 1 have been assigned higher
ranks than in Condition 2. In Table 7.3 when the totals of the ranks are
added up separately for each condition the total of ranks is higher in
Condition 1 than in Condition 2.
These links between higher scores, higher means, higher ranks and
higher rank totals in Condition 1 demonstrate the connections between
scores and ranks for assessing ordinal data.
7.4
We will now take you through the answers you need to give in the Decision
Chart (inside back cover) in order to select an appropriate statistical test
for analysing ordinal data for two conditions in which two groups of
different participant do each condition.
The answer is Yes because the aim of the experiment is to test
predicted differences between experimental conditions.
Table 7.3 Rank totals for the two conditions
Ranks Ranks
5.5 1
7 2
8 3
4 5.5
Rank totals 24.5 11.5
The answer is two conditions because the experiment is testing dif-
ferences between two conditions.
The answer is unrelated design because different participants are
doing each condition.
If you proceed down the Decision Chart in this way, you will arrive
at the MannWhitney test (unrelated). The MannWhitney (unrelated)
is the appropriate non-parametric test for an unrelated design with two
conditions.
Section 7.5 will provide practical advice about what you have learned so
far in this chapter about how to rank overall scores and how to calculate
rank totals between conditions.
Using the MannWhitney test (unrelated)
7.5
When to use
The MannWhitney test (unrelated) should be used for an unrelated
design when different participants are doing two conditions and the data
are ordinal.
Research hypothesis
The prediction was that a group of participants who are given a relevant
title before learning a text would recall the text better than a group of
participants given no title.
?
Why is it better to use different participants in each condition to test this Question 7.2
hypothesis?
Sample data
One group of participants was given a relevant title before reading a text
(Condition 2). A second group of participants was given no title before
reading the text (Condition 1). Participants in both groups were tested for
their memory of the text. The results are shown in Table 7.4.
Notes on Table 7.4
1 The overall ranks 1 and 2 are the result of an overall ranking of all
scores.
2 The rank totals (T
1
and T
2
) have been added up separately for each
condition,
3 Note that there are different numbers of participants in each group,
6 participants in Condition 1 and 8 participants in Condition 2.
Rationale
The aim of the MannWhitney test (unrelated) is to compare the rank
totals in the two conditions based on an overall ranking. If the
ranked differences between conditions are random, as stated by the null
hypothesis, there should be roughly equal ranks in the two conditions. If
there is a large preponderance of low or high ranks in one condition in
the predicted direction, the differences between conditions well justify the
rejection of the null hypothesis.
Table 7.4 Recall scores for texts with and without a title
Condition 1
(no title)
Overall ranks
(ranks 1)
Condition 2
(title)
Overall ranks
(ranks 2)
3 3 9 13
4 4 7 10.5
2 1.5 5 6
6 8.5 10 14
2 1.5 6 8.5
5 6 8 12
5 6
7 10.5
Rank totals T
1
= 24.5 T
2
= 80.5
Means 3.67 7.13
Step-by-step instructions for calculating the value of U
These are given in Box B.
The formula for U in step 3 may look rather daunting. All that is
happening is that U has to take into account the unequal number of
participants in each group (n
1
and n
2
). The convention is to subtract T
1
from the formula.
Box B
Step-by-step instructions for calculating the value of U
Rank all the scores (taking both 1
groups together as an overall set of
ranks), giving rank 1 to the lowest
score in the usual way.
See overall ranks for both conditions
taken together in Table 7.4.
After ranking all the scores, add the 2
rank totals for each condition.
T
1
= 24.5
T
2
= 80.5
Calculate U using the formula: 3
U = n
1
n
2
+
n
1
(n
1
+ 1)
2
T
1
U = 6 8 +
6(6 + 1)
2
24.5
= 48 +
42
2
24.5
= 48 + 21 24.5
= 44.5
n
1
= number of participants in
Condition 1
n
1
= 6
n
2
= number of participants in
Condition 2
n
2
= 8
T
1
= rank total for Condition 1 T
1
= 24.5
Next substitute the values of U, n
1
and 4
n
2
in the following formula and
calculate U:
U = n
1
n
2
U U = 6 8 44.5
= 3.5
Take the smaller value of U or U as U. 5 U = 3.5
Looking up the signicance of U in Table B
Tables B(1) and B(2) give the probabilities for testing one-tailed and two-
tailed hypotheses. It may seem odd to have to look up U in two separate
tables. The reason is that the tables provide the U values for different
signicance levels. Table B(1) gives the probabilities for a one-tailed
hypothesis at p < 0.05. Table B(2) gives the probabilities for p < 0.01 (one-
tailed) and p < 0.02 (two-tailed). The smaller value of U has to be equal to
or less than the values in Table B.
Looking at Table B(1) rst, the number of participants in Condition 1
(n
1
= 6) is listed along the top row and the number of participants in
Condition 2 (n
2
= 8) is listed in the left-hand column. Locate 6 along the
top and 8 down the column. Where these two meet the value of U is 10.
Our calculated U of 3.5 is smaller than this so the null hypothesis can be
rejected. Looking at Table B(2) the U of 3.5 is also smaller than the value
of 6 for n
1
= 6 and n
2
= 8. The null hypothesis for a one-tailed hypothesis
can be rejected (p < 0.01).
Conclusions
It is essential to refer back to the means in Table 7.4 to check whether
the results are signicant in the predicted direction. The means show a
difference in favour of Condition 2 (title). It can be concluded that the
prediction that participants will recall more of a text when given an
appropriate title was supported (p < 0.01).
Progress box seven

In unrelated designs the scores in the two conditions are produced by
different participants.
Individual scores are assigned an overall ranking across both
conditions.
Totals of ranks are added separately for each condition.
The appropriate statistical test for analysing unrelated ordinal data
for two conditions is the MannWhitney test (unrelated).
?
An experimenter predicted that words presented on the left-hand side of Question 7.3
a screen would be recognized more quickly because of the direction of
reading left to right in English compared with words presented on the
right-hand side. The measure was the time taken to recognize the words.
The data are given in Table 7.5.
(a) Following steps 1 and 2 in Box B, rst calculate the overall ranks of
scores in Table 7.5 and add them on the overall ranks columns.
(b) Add the rank totals separately for each condition.
(c) Do the rank totals support the research hypothesis? (Remember that
lower rank totals represent quicker recognition times.)
Table 7.5 Data for MannWhitney (unrelated)
Left-hand
presentation
Overall ranks
(ranks 1)
Right-hand
presentation
Overall ranks
(ranks 2)
9 8
3 5
2 7
6 20
1 11
5 6
Rank totals T
1
= T
2
=
Introduction to parametric t tests 8
Comparisons between parametric and non-parametric tests
8.1
The parametric t tests, which we will be introducing in this chapter, are
designed to test differences between two conditions. In this way, they are
like the non-parametric tests for two conditions that were discussed in
Chapters 6 and 7. In other ways, parametric tests and non-parametric tests
are quite different.
We will start by emphasizing the similarities. In both cases the tests
analyse differences between two experimental conditions. The research
hypothesis predicts that there will be differences between the two
experimental conditions.
Another similarity between parametric t tests and non-parametric tests
is that they use similar methods for analysing data from same participants
and data from groups of different participants.
Chapter 6 described the way in which the non-parametric Wilcoxon test
(related) deals with data from a related design, when the same participants
do both experimental conditions. Participants who do both experimental
conditions produce pairs of scores for the two conditions. This means that
for each participant it is possible to calculate the differences between
scores in the two conditions.
The parametric t test (related) also compares differences between pairs
of scores on two conditions for each participant.
Chapter 7 described the way in which the non-parametric Mann
Whitney test (unrelated) deals with data from an unrelated design, when
groups of different participants do each of the experimental conditions.
Because a different group of participants does each of the conditions, it is
not possible to compare differences between scores. Instead all the scores
have to be dealt with as individual scores.
The parametric t test (unrelated) also assumes that there is no basis for
comparing scores directly. The aim is to investigate whether the mean of
the scores for the group of participants doing one condition is higher than
the mean of the scores for the different group of participants who are
doing the other condition.
The major differences between the two types of statistical tests arise
from the different way in which the data are measured.
Non-parametric tests are based on ordinal data in which scores can be
ordered from the lowest score to the highest score. Scores can be ranked,
giving the lowest rank to the lowest score and so on until the highest rank
is assigned to the highest score.
Parametric t tests are based on interval data. Interval data are so called
because the intervals between scores are considered to be equal. This
makes it possible to carry out numerical calculations instead of simply
ranking the scores.
Numerical calculations for parametric tests
8.2
One of the most common numerical calculations for parametric tests is to
square numbers and to add the squared numbers in various combinations.
Squaring a number simply means multiplying a number by itself. You
will probably remember from your school days that when you calculate the
squares of positive and minus scores you always end up with positive
numbers. For example:
3 3 = +9
3 3 = +9
?
Calculate the squares of the following numbers. Question 8.1
(a) 5
(b) 5
(c) 1
(d) 3
The next step is to add up squared numbers to arrive at a total of the
squared numbers. Another word for total is sum. The term used is sum
of squares (SS). The symbol for sum is . This symbol is the Greek letter
for S (pronounced sigma).
It is a general principle for parametric tests that variance is calculated by
squaring numbers and summing them in various combinations. You will
be given detailed step-by-step instructions for each parametric test about
which numbers should be squared and summed.
Calculating variances
8.3
Based on the numerical calculations that can be carried out on interval
data, it is possible to calculate precise amounts of variability in par-
ticipants scores. The calculated differences in scores between the two
conditions can be compared with the total amount of variability in scores.
When using parametric t tests the usual term for describing variability is
variance. Variance represents calculated estimates of variability in scores.
Predicted variance represents the predicted differences in scores between
the two conditions.
Total variance represents all the variability in scores, including the
variance due to predicted differences and irrelevant differences between
individual participants.
Ratio of variances
8.4
Another way of expressing the comparison between predicted variance
and total variance is as a ratio between the two types of variance.
The rst type of variance is the predicted variance. The prediction is
that there will be differences between two conditions as a result of the
independent variable selected by the researcher.
The second type of variance is total variance. This is the result of all
the variables affecting participants performance. Total variance includes
the variance caused by the predicted independent variable, as well as
unpredicted variance due to individual differences between participants
and the effects of irrelevant variables.
The comparison between these two types of variance can be expressed
as a ratio:
Predicted variance due to independent variable
Total variance due to all variables
Parametric t tests test whether the predicted variance is large enough to
result in a high ratio of variance in which the predicted variance between
two conditions is relatively large in comparison with the total variance.
As we shall see, it is the concept of a ratio between predicted and total
variance that underlies all parametric statistical tests. You do not need to
worry too much about this now, but you will come across this ratio when
the t tests are described. Full details about the implications of predicted
and total variance will be given in the description of th t tests in the next
two chapters.
Progress box eight

Parametric t tests are similar to non-parametric tests for testing
differences between two experimental conditions.
Both parametric t tests and non-parametric tests distinguish between
related and unrelated designs.
Parametric tests are different from non-parametric tests because
they require interval data, which provide the basis for numerical
calculations.
Ratios of variances compare predicted variance with total variance.
t test (related) 9
9.1
In related designs the same participants do both conditions. Each par-
ticipant produces a pair of scores, one score for Condition 1 and one score
for Condition 2.
This makes it possible to make a direct comparison between each
participants scores in the two conditions. Each participant has a pair of
related scores in both conditions so the differences between the scores can
be calculated for each participant.
Because the t test (related) is based on pairs of related scores, another
name for the t test (related) is the paired t test.
Squaring differences
9.2
You should remember from Chapter 8 that one of the most common
numerical calculations used for interval data is to square numbers and add
them to produce sums of squares. For related designs it is the differences
between conditions that are squared.
Table 9.1 shows the plus and minus differences between the scores on
Table 9.1 Squares of differences between conditions
Differences
d (1 2)
Squares of
differences (d
2
)
1 10 5 +5 25
2 8 7 +1 1
3 7 10 3 9
4 3 8 5 25
t test (related) 49
two conditions, obtained by subtracting the scores for Condition 2 from
the scores in Condition 1.
For the t test (related) each plus or minus difference is squared. The
squares of both the plus or minus differences are always a positive number.
The next step is to nd the totals of the scores, the differences and the
sum of squared differences, as shown in Table 9.2.
?
Using the totals for each condition, calculate the mean for each condition Question 9.1
in Table 9.2.
Note that the total of the differences takes into account the plus and minus
differences (plus 6 8 = 2).
The sum of positive squares always results in a positive number
(here 60).
Ratio of variances
9.3
The predicted variance is measured as the sum of squared differences
between the scores in the two conditions. The research hypothesis predicts
differences between two experimental conditions as a result of the
independent variable selected by the experimenter.
The predicted variance has to be compared with the total variance in
scores. Total variance is based on all the scores produced by participants.
Total variance is measured using the sum of squared differences between
the two conditions.
The t test (related) calculates a ratio between the predicted variance and
the total variance.
Table 9.2 Totals (sums) of differences and squared differences
Differences
d (1 2)
Squares of
differences (d
2
)
1 10 5 +5 25
2 8 7 +1 1
3 7 10 3 9
4 3 8 5 25
Totals (sums) 28 30 2 60
9.4
The answer to this question is Yes because the aim of the experi-
ment is to test predicted differences between experimental conditions.
both conditions.
The answer is parametric because parametric statistical tests are
suitable for analysing numerical interval data.
You will nd that if you answered the above four questions in the Decision
Chart (inside back cover) as suggested, you will arrive at the t test (related).
The t test (related) is the appropriate parametric test for a related design
with two conditions.
Using the t test (related)
9.5
When to use
The t test (related) should be used for a related design when the same
participants are doing two conditions and the data are interval.
Research hypothesis
The experimenter predicted that participants would recall more words
from a simple text than from a complex text.
?
Is this a one-tailed or a two-tailed hypothesis? Question 9.2
t test (related) 51
Sample data
All participants were given 2 minutes to read a simple text and a
complex text and after 10 minutes they were asked to recall the words.
To counterbalance possible order effects of doing both conditions,
half of the participants were presented with the simple text rst and
the other half with the complex text rst. The results are shown in
Table 9.3.
Notes on Table 9.3
1 The differences (d) are calculated by subtracting the scores in Condition
2 from the scores in Condition 1.
2 The sum of differences (
d) is calculated taking into account the plus

and minus signs.
3 The sum of the squares of differences (
d
2
) is calculated by adding the
positive squares.
Rationale
When the same participants are used for both conditions, the predicted
variance is calculated as the sum of the squared differences between the
scores in both conditions. The predicted variance between conditions is
expressed as a proportion of total variance. If there are only random
differences between the scores in the two conditions, as stated by the
Table 9.3 Recall scores for simple and complex texts
Participants
Condition 1
(simple text)
Condition 2
(complex text) d d
2
1 10 2 8 64
2 5 1 4 16
3 6 7 1 1
4 3 4 1 1
5 9 4 5 25
6 8 5 3 9
7 7 2 5 25
8 5 5 0 0
9 6 3 3 9
10 5 4 1 1
d = 27

d
2
= 151
Means 6.4 3.7
null hypothesis, the variance due to the predicted differences would be
relatively small in relation to the total variance in scores. In this case the
null hypothesis cannot be rejected.
Step-by-step instructions for calculating t (related)
These are given in Box C.
Box C
Step-by-step instructions for calculating t (related)
Calculate the differences between 1
participants scores by subtracting
Condition 2 scores from Condition 1.
See column d in Table 9.3.
Sum differences, taking into account 2
pluses and minuses.
d = 27
Square the differences. 3 See column d
2
in Table 9.3.
Sum the squared differences. 4

d
2
= 151
Square the total of differences. 5 (
d
2
) = 27 27 = 729
Find the total number of participants. 6 N = 10
Find t from the formula: 7
t =
d
2
(
d)
2
N 1
t =
27
10 151 (27)
2
10 1
=
27
1510 729
9
=
27
86.78
= take the square root
=
27
9.315
= 2.89
Calculate the degrees of freedom. 8 df = N 1 = 10 1 = 9
t test (related) 53
Notes on calculations in Box C
1 Look carefully at the denitions of symbols in Box C and follow the
calculations on the right-hand side of the box. Note that you also have
to calculate the square of the total differences (27 27 = 729).
2 In the main formula in step 7 the sum of the differences between pairs of
scores (
d) represents the predicted differences between the conditions

(on the top line of the formula).
3 The equation on the bottom line of the formula represents total
variance in terms of sums of squares. The whole formula represents a
ratio between predicted variance and total variance.
4 In step 8 you are asked to calculate the degrees of freedom (df ). Take
it on trust that df equals the number of participants (N) minus 1. In our
example df = 10 1 = 9. We will be explaining the principles of degrees
of freedom in a later chapter.
5 The calculations in Box C will be much easier if you invest in a calc-
ulator which has a sum of squares button (
x
2
) which automatically
works out squares and adds them up. If you have a calculator with
memory you can put each square into memory and record the total
sum of squares. A square root button is essential to complete the last
instruction to take the square root.
Looking up the signicance of the t test (related) in Table C
Table C is the appropriate table for the t test for two conditions, whether
the t test is related or unrelated. Table C gives the probabilities for the
values of t. The calculated t has to be equal to or larger than the values in
Table C. You need to use the df (in our example df = 10 1 = 9).
To nd the value of t look along the row for df = 9. The value of t for
(p < 0.05) for a one-tailed hypothesis is 1.833. Our calculated t of 2.89 in
Box C is larger than this value. In fact, looking along the df = 9 row the
calculated t is larger than the value of 2.821 for (p < 0.01) for a one-tailed
hypothesis. The null hypothesis can be rejected (p < 0.01)
Conclusions
It is essential to look at the means in Table 9.3 to check whether any
signicant differences are in the predicted direction. The means in
Table 9.3 show that the results are signicant in the predicted direction,
supporting the one-tailed hypothesis that more words are recalled from a
simple text than from a complex text (p < 0.01).
Progress box nine

In related designs pairs of scores in the two conditions are produced
by the same participants.
The total of differences is calculated between the pairs of scores for
each participant. This represents predicted variance.
In order to calculate total variance the squares of differences are
added to produce a sum of squares.
The appropriate parametric statistical test for calculating a ratio
between predicted and total variance when analysing related interval
data for two conditions is the t test (related).
?
Question 9.3
(a) Using the data in Table 9.4, calculate the means for the two conditions.
(b) Calculate the plus and minus differences (d) between the two con-
ditions (Condition 1 Condition 2) and insert these in the d column
in Table 9.4.
(c) Square the differences and insert these in the d
2
column.
(d) Calculate the sum of differences (
d), taking into account plus and

minus differences.
(e) Calculate the sum of squares (
d
2
).
Table 9.4 Data for t test (related)
Participants Condition 1 Condition 2 d d
2
1 6 2
2 7 1
3 8 3
4 10 4
5 8 3
6 8 2
7 5 7
8 3 4
Means

d =

d
2
=
t test (unrelated) 10
10.1
In unrelated designs different groups of participants do each condition.
Because different participants do only one of the experimental conditions,
the scores in the two conditions are unrelated. It is not possible to compare
scores in both conditions for each participant.
Instead all scores have to be dealt with as indvidual scores. The t test
(unrelated) is based on comparing the means for the two groups doing
each condition. This is because there is no basis for comparing differences
between related pairs of scores for each participant.
Because the t test (unrelated) is based on unrelated scores for two con-
ditions, which are independent of each other, another name for the t test
(unrelated) is the independent t test
Squaring scores
10.2
You should remember from Chapter 8 that one of the most common
numerical calculations used for parametric tests is to square numbers and
add them to produce sums of squares.
For unrelated designs all the individual scores are squared. An example
is given in Table 10.1.
To make it clear that different groups are doing the two con-
ditions, Table 10.1 shows Group 1 doing Condition 1 and Group 2 doing
Condition 2.
?
Do the scores of 5 in Condition 1 and Condition 2 represent tied scores? Question 10.1
The next step is to add the totals of the scores and the sums of squares
separately for Condition 1 and Condition 2. These totals of scores and
squares are shown in Table 10.2.
Ratio of variances
10.3
The predicted variance is based on the difference between the means for
the two conditions. The research hypothesis predicts differences between
two experimental conditions as a result of the independent variable
selected by the experimenter.
The predicted variance has to be compared with the total variance in
scores. Total variance is based on all the scores produced by participants.
Total variance is measured using the sums of squares of scores in each of
the two conditions.
The t test (unrelated) calculates a ratio between the predicted variance
and the total variance.
Table 10.1 Squares of individual scores
Group 1 Group 2
Scores Squares Scores Squares
2 4 4 16
5 25 5 25
4 16 3 9
3 9 6 36
Table 10.2 Sums of scores and squares
Group 1 Group 2
Scores Squares Scores Squares
2 4 4 16
5 25 5 25
4 16 3 9
3 9 6 36
Totals (sums) 14 54 18 86
Means 3.5 4.5
10.4
The answer is parametric because parametric statistical tests are
suitable for analysing numerical interval data.
Chart (inside back cover) as suggested, you will arrive at the t test
(unrelated). The t test (unrelated) is the appropriate parametric test for an
unrelated design with two conditions.
Using the t test (related)
10.5
When to use
The t test (unrelated) should be used for an unrelated design when different
participants are doing two conditions and the data are interval.
Research hypothesis
The experimenter predicted that more words would be recalled from a
simple text than from a complex text.
Sample data
Two groups of different participants were used for each condition. One
group was asked to read a simple text and the other group of different
participants was given a complex text. Both groups were given 2 minutes
to read the text and after 10 minutes were asked to recall the words. The
results are shown in Table 10.3.
Notes on Table 10.3
1 The means are calculated for Condition 1 and Condition 2.
2 The individual scores are squared.
3 The sums of squares are summed separately for each group of
participants.
Rationale
When different participants are used for each condition, the predicted
variance is calculated as the difference between the means in the two
conditions. The predicted variance is expressed as a ratio of total variance
in scores. If there are only random differences between the scores in the
two conditions, as stated by the null hypothesis, the variance due to
the predicted differences between conditions would be relatively small in
relation to the total variance in scores. In this case the null hypothesis
cannot be rejected.
Step-by-step instructions for calculating t (unrelated)
These are given in Box D.
Table 10.3 Recall scores for simple and complex texts
Condition 1 (simple text) Condition 2 (complex text)
Scores Squared scores Scores Squared scores
(x
1
) (x
1
2
) (x
2
) (x
2
2
)
10 100 2 4
5 25 1 1
6 36 7 49
3 9 4 16
9 81 4 16
8 64 5 25
7 49 2 4
5 25 5 25
6 36 3 9
5 25 4 16
x
1
= 64

x
2
1
= 450

x
2
= 37

x
2
2
= 165
Means M
1
= 6.4 M
2
= 3.7
Box D
Step-by-step instructions for calculating t (unrelated)
Sum the totals of scores for each condition (see 1
Table 10.3) and square these.
(
x
1
)
2
= 64 64 = 4096
(
x
2
)
2
= 37 37 = 1369
Calculate the means for each condition. 2 M
1
= 6.4
M
2
= 3.7
Square the individual scores for each condition. 3 See columns for squared
scores in Table 10.3.
Sum the squared scores for each condition separately. 4

x
2
1
= 450
x
2
2
= 165
Number of participants in each condition 5 n
1
= 10
n
2
= 10
Calculate the value of t using the formula 6
t =
M
1
M
2
x
2
1

(
x
1
)
2
n
1
x
2
2

(
x
2
)
2
n
2
(n
1
1) + (n
2
1)

1
n
1
+
1
n
2
t =
6.4 3.7
450
4096
10

+
165
1369
10

9 + 9

1
10
+
1
10
=
2.7
(450 409.6) + (165 136.9)

18

1
5
=
2.7
3.806 0.2
=
2.7
0.7612
= 3.096
(N.B. It does not matter whether subtracting M
2
from M
1
results in a plus or minus
number. When looking up the value of t simply ignore the sign)
Calculate the degrees of freedom by subtracting one 7
from the number of participants in each condition
and adding together the results.
n
1
1 = 9
n
2
1 = 9
df = 9 + 9 = 18
Notes on calculations in Box D
1 Look carefully at the denitions of symbols in Box D and follow the
calculations on the right-hand side of the box.
2 In the main formula in step 6 the difference between the means (M
1

M
2
) represents the predicted differences between the conditions (on the
top line of the formula).
3 The equation on the bottom line of the formula represents total
variance in terms of sums of squares. The reason for the horrendous-
looking equation on the bottom line is that variances have to be cal-
culated separately for each group of participants and then added to give
total variance. All this will seem less formidable (we hope!) as you work
your way through the formula. The whole formula represents a ratio
between predicted variance and total variance.
4 In step 7 you are asked to calculate the degrees of freedom (df ). The
special feature of the df for unrelated designs is that you have to subtract
1 from the number of participants in each group (n
1
and n
2
). If there are
10 participants in each group then the df for n
1
are 10 1 = 9. The
df for n
2
are 10 1 = 9. Adding these two together, df = 18. Follow
the instructions in Box D. You will be given full details about calculating
the df in a later chapter.
5 It is possible to do these calculations on a calculator with a square root
button and a memory store. There are also computer programs that will
carry out all the calculations automatically, including signicance levels.
But it is worth following the steps in Box D in order to understand the
rationale for the t test (unrelated).
Looking up the signicance of the t test (unrelated) in Table C
Table C is the appropriate table for the t test for two conditions, whether
the t test is related or unrelated. Table C gives the probabilities for the
values of t. The calculated t has to be equal to or larger than the values
in Table C. You need to use the df (in our example df = n
1
1 + n
2

1 = 18).
To nd the value of t look along the row for df = 18. The value of t for
p < 0.05 for a one-tailed hypothesis is 1.734. Our calculated t of 3.096 in
Box D is larger than this value. In fact, looking along the df = 18 row, the
calculated t is larger than the critical value of 2.878 for p < 0.005. This is a
very small probability of random of random differences in scores. The null
hypothesis can be rejected (p < 0.005).
Conclusions
It is essential to look at the means in Table 10.3 to check whether any
signicant differences are in the predicted direction. This is especially
important if a computer program has been used. The means in Table 10.3
show that the results are signicant in the predicted direction, supporting
the one-tailed hypothesis that more words are recalled from a simple text
than from a complex text (p < 0.005).
Progress box ten

In unrelated designs the scores in the two conditions are produced by
Means are calculated for the two conditions. This represents pre-
dicted variance.
In order to calculate total variance the squares of scores for each
condition are added to produce sums of squares.
The appropriate parametric statistical test for calculating a ratio
between predicted and total variance when analysing unrelated
interval data for two conditions is the t test (unrelated).
Friedman test (related) 11
Scores from same participants for three conditions
11.1
We will now be returning to non-parametric tests, this time for three
(or more) conditions. The Friedman test is for related designs in which
the same participants do all the experimental conditions. As with the
statistical tests for two conditions, we are introducing rst the simpler non-
parametric tests for dealing with three conditions.
The Friedman test (related) is used when the same participants do all
three conditions. Each participant produces three scores, one in each of
the three conditions. This makes it possible to make a direct comparison
between each participants three scores in all the conditions.
The scores in the conditions are related. But, because there are three
conditions it is not possible to calculate the differences between two con-
ditions, as we did for the two conditions in the Wilcoxon test (related) and
the t test (related). A different kind of ranking has be carried out for three
conditions.
An example is given in Table 11.1 of an experiment in which partici-
pants are presented with three conditions: lists of very rare words (Con-
dition 1), fairly common words (Condition 2) and very common words
Table 11.1 Recall scores for different types of word lists
Participants
Condition 1
(rare words)
Condition 2
(common words)
Condition 3 (very
common words)
1 2 3 4
2 3 5 7
3 1 6 4
4 2 6 7
(Condition 3). Because the conditions are done by the same participants
there are three scores for each participant.
Ranking three conditions
11.2
The next step is to rank the scores in Table 11.1. The aim of ranking
ordinal data is to test whether participants have higher scores in some
conditions than in other conditions. If you look at the scores in Table 11.1
for three conditions you will realize that it is not possible to calculate
differences between pairs of scores. But we still need to be able to compare
the three scores produced by each participant.
The way the three scores in each of the conditions are ranked may seem
rather bizarre but there is a good rationale for it. The three scores for each
participant are ranked horizontally across the three related conditions.
The aim of ranking the three scores for each participant is that it provides
information about whether each participant produced higher or lower
scores in the three related conditions.
The three conditions in Table 11.1 are repeated in Table 11.2 but with
horizontal ranks for the three related scores in the three conditions for
each participant added in the Ranks columns.
Notes on Table 11.2
1 The top horizontal row of ranks for participant 1 shows that the score
of 2 in Condition 1 is the lowest of the three scores and so is assigned a
rank of 1. The score of 3 in Condition 2 is the next lowest score and gets
a rank of 2. The score of 4 in Condition 3 is the highest of the three
scores for participant 1 and so is assigned the highest rank of 3.
2 If you look at the scores and ranks in Table 11.2 for Participant 2, the
scores and ranks are in the same rank order as for Participant 2.
3 Participant 3 is slightly the odd one out. The lowest score of 1 and rank
1 was for Condition 1. But this participant obtained a higher recall
score of 7 for Condition 2 compared with a score of 4 for Condition 3.
Table 11.2 Horizontal ranks of three related scores
Condition 1 Condition 2 Condition 3
Participants Scores Ranks Scores Ranks Scores Ranks
1 2 1 3 2 4 3
2 3 1 5 2 7 3
3 1 1 7 3 4 2
4 2 1 6 2 7 3
This participant recalled more words from the list of fairly common
words in Condition 2 than the very common words in Condition 3. So
the ranks for Participant 3 are in the order 1,3,2.
Assigning tied ranks
11.3
As is always the case when ranking ordinal data, sometimes there will be
identical scores. These are dealt with in the normal way by assigning the
average of the ranks they would have been entitled to. An example is given
in Table 11.3.
Look at the top row for Participant 1. There are two scores of 2. Tied
ranks of 1.5 are allocated to these two scores. The score of 4 in Condition
3 is assigned the next available rank of 3.
?
Insert the three horizontal ranks for the related scores for Participants 5 Question 11.1
and 6 in the Ranks columns.
Table 11.3 Tied ranks
1 2 1.5 2 1.5 4 3
2 3 1 5 2 7 3
3 1 1 7 3 4 2
4
5
6
2
3
3
1
7
3
6
5
6
7 3
5 3 7 5
6 3 3 6
Rank totals
11.4
After having ranked the three related scores horizontally for the three
conditions, the next step is to add up the totals of ranks separately for each
condition. The idea is to test whether there is a preponderance of higher or
lower ranks in any of the three conditions. Taking the data in Table 11.3,
the total of ranks for each condition is shown in Table 11.4.
The ranks have been added separately for each condition to produce a
total of ranks for each condition. The rank totals are 4.5 for Condition 1,
8.5 for Condition 2 and 11 for Condition 3. As predicted, fewer very rare
words were recalled in Condition 1 and most words recalled for the very
common words in Condition 3.
?
Question 11.2
(a) Use the totals of scores for each condition in Table 11.4 to calculate
the means for each of the three conditions.
(b) Do the means support the prediction that more very common words
are recalled (Condition 3) than very rare words (Condition 1)?
The Friedman test (related) can deal with three or more conditions. If
there were four conditions there would be four related scores to be ranked
horizontally for each participant. Four related scores would be allocated
four horizontal ranks, from rank 1 for the lowest score to rank 4 for the
highest score.
Table 11.4 Rank totals
Condition 1
(rare words)
Condition 2
(common words)
Condition 3 (very
common words)
1 2 1.5 2 1.5 4 3
2 3 1 5 2 7 3
3 1 1 7 3 4 2
4 2 1 6 2 7 3
Totals 8 4.5 20 8.5 22 11
11.5
The answer is three or more conditions because the experiment is
testing differences between three (or more) conditions.
both conditions.
Chart (inside back cover) as suggested, you will arrive at the Friedman test
(related). The Friedman test (related) is the appropriate non-parametric
test for a related design with three or more conditions.
Using the Friedman test (related)
11.6
When to use
The Friedman test (related) should be used for a related design when the
same participants are doing three or more conditions and the data are
ordinal.
Research hypothesis
The experimenter predicted that children would rate some illustrations as
being more appealing than other illustrations.
Sample data
A publisher producing a series of childrens books wants to choose from
three types of illustrations the one which is most appealing to children.
Eight children are asked to rate all three illustrations on a ve-point
scale. The children were presented with the three illustrations in dif-
ferent orders. The rating scores for the three illustrations are shown in
Table 11.5.
Notes on Table 11.5
1 Each row of ranks represents the horizontal ranks for the three related
scores for the three conditions for each participant.
2 You will notice that some of the ranked scores are the same and are
assigned the average of the tied ranks they would have been entitled to.
3 The rank totals for each condition are added to produce rank totals for
each of the three conditions.
Rationale
The aim of Friedman test (related) is to test whether the rank totals are
different for the three conditions. If there are random differences between
the rank totals for all three conditions, as stated by the null hypothesis, the
rank totals would be approximately equal, because there would equal
numbers of low and high ranks in each of the conditions. If the rank totals
show a preponderance of high ranks or low ranks in one or other of the
conditions, the null hypothesis can be rejected.
Table 11.5 Rating scores for three illustrations
Condition 1
(illustration A)
Condition 2
(illustration B)
Condition 3
(illustration C)
Participants
Rating
scores Ranks
Rating
scores Ranks
Rating
scores Ranks
1 2 1 5 3 4 2
2 1 1 5 3 3 2
3 3 1 5 2.5 5 2.5
4 3 2 5 3 2 1
5 2 1 3 2 5 3
6 1 1 4 2.5 4 2.5
7 5 3 3 2 2 1
8 1 1 4 3 3 2
Rank totals 11 21 16
Means 2.25 4.25 3.50
Step-by-step instructions for calculating Friedman test (related)
These are given in Box E.
Box E
Step-by-step instructions for calculating the value of
2
r
Rank the scores for each separate 1
participant horizontally across each
row, giving 1 to the smallest score, 2 to
the next score and 3 to the highest score.
See Ranks columns in Table 11.5
allocating ranks 1, 2, 3 for each
participant (sometimes tied, e.g. 2.5
for participants 3 and 6).
Calculate the rank totals for each 2
condition.
T
1
= 11, T
2
= 21, T
3
= 16
Note the following symbols: 3
number of conditions (C) C = 3
number of participants (N) N = 8
T
2
= squares of rank totals T
2
1
= 11 11 = 121
T
2
2
= 21 21 = 441
T
2
3
= 16 16 = 256
T
2
= 11
2
+ 21
2
+ 16
2
= 818
Calculate the value of
2
r
using the 4
formula
2
r
=
12
NC(C + 1)
T
2

2
r
=
12
(8 3)(3 + 1)
(11
2
+ 21
2
+ 16
2
)
3N(C + 1) 3 8(3 + 1)
=
12
24 4
(121 + 441 + 256)
(24 4)
=
12
96
818
96
= 102.25 96
= 6.25
Calculate the degrees of freedom: the 5
number of conditions minus one.
df = C 1
= 3 1
= 2
Notes on calculations in Box E
1 The Friedman test (related) is a non-parametric test based on ordinal
data, in which three scores are ranked horizontally.
2 The test statistic is denoted by
2
r
, which is pronounced ky r squared.
is the Greek letter chi.
3 The calculations involve calculating the squares of the three rank totals
and summing them to produce a sum of squares.
4 The formula in step 4 has to take into account the number of conditions
and the squared rank totals.
5 You are told to calculate the degrees of freedom (df). The df are the
number of conditions, C, minus 1 (in our example 3 1 = 2). Note that
you only need the df for experiments when there are relatively large
numbers of participants doing all three conditions.
Looking up the signicance of
2
r
in Table D or E
Note that there are quite often special tables for small numbers of
conditions and participants and other tables for larger numbers.
Table D: up to nine participants
Table D covers three conditions for up to nine participants doing three
conditions. Another point to notice about Table D is that it gives a list of
all possible probabilities that can be read off for any calculated
2
r
. The
only relevant percentage values in Table D are those which achieve con-
ventional levels of signicance (p < 0.05). These are found at the bottom
of the lists of probabilities.
The calculated
2
r
has to be equal to or larger than the values in the
table. Our calculated
2
r
in Box E is 6.25 and there were 8 participants.
Look down the column for N = 8 in Table D. For the calculated 6.25 the
probability is 0.047. This probability is less than (p < 0.05) so the null
Table E: more than nine participants
If there were more than nine participants it is possible to consult a more
general statistical table, known as chi-square (
2
) in Table E. To look up
probabilities in Table E the df is the number of conditions, C, minus 1. In
our example the df are 3 1 = 2. In Table E look along the df = 2 row. The
calculated 6.25 is larger than the value of 5.99 for p< 0.05. The mull
Conclusions
It is essential to look at the means in Table 11.5 to check which of the
illustrations was preferred by the children. It appears from the means in
Table 11.5 that they preferred illustration B, since this received the highest
ratings, with illustration C next and illustration A last. The Friedman
test (related) can only test whether there are overall differences between
conditions, which is the equivalent of a two-tailed hypothesis.
Progress box eleven

In related designs three related scores in three conditions are pro-
duced by the same participants.
The three related scores for each participant are ranked horizontally.
The totals of ranks are added for each condition.
The appropriate non-parametric statistical test for analysing
related ordinal data for three conditions is the Friedman test
(related).
?
Five participants were given three passages of prose to read, one typed Question 11.3
in black print, one in red print and one in green print. Scores were the
number of ideas recalled, as shown in Table 11.6.
(a) Using the data in Table 11.6, calculate the mean scores for the three
conditions.
(b) What is the independent variable and what is the dependent
variable?
(c) What is the appropriate statistical test and why?
Table 11.6 Data for Friedman (related)
Participants
Condition 1
(black print)
Condition 2
(red print)
Condition 3
(green print)
1 4 5 6
2 2 7 7
3 6 6 8
4 3 7 5
5 3 8 9
(d) Following the step-by-step instructions in Box E, assign horizontal
ranks to the three scores for each participant.
(e) Add the rank totals for each condition.
KruskalWallis test (unrelated) 12
Scores from different participants for three
12.1
(or more) conditions
The KruskalWallis test is a non-parametric test based on ranking ordinal
data. As with the statistical tests for two conditions, we are introducing
rst the simpler non-parametric tests for dealing with three conditions.
In unrelated designs different participants do each condition. Because
different participants do only one of the conditions, it is not possible to
compares scores in three conditions for each participant. Instead all the
individual scores have to be given overall rankings across all three
conditions
An example is given in Table 12.1 for an experiment in which different
groups of participants are presented with one each of three conditions.
These conditions are very rare words (Group 1 doing Condition 1), fairly
common words (Group 2 doing Condition 2) and very common words
(Group 3 doing Condition 3).
Table 12.1 Recall scores for different types of word lists
Group 1
(rare words)
Group 2
(common words)
Group 3
(very common words)
Scores Ranks Scores Ranks Scores Ranks
2 3 11
3 5 7
1 6 4
0 6 8
Overall ranking of scores
12.2
You should remember from Chapter 7, which described the Mann
Whitney test (unrelated), that when there is no basis for direct com-
parisons, all the individual scores are ranked overall across all three
conditions.
Exactly the same thing happens with the KruskalWallis test
(unrelated). The individual scores in all three conditions are assigned
overall ranks.
In Table 12.2 overall ranks have been assigned to the individual scores in
all three condition in the Ranks columns.
We will now take you through the assignment of overall ranks of all
the scores across all three conditions, just as was done for two unrelated
conditions for the MannWhitney (unrelated).
One participant in Condition 1 scored none of the very rare words
correctly, which is a score of 0. This is the lowest score so it is assigned the
lowest rank of 1. The next lowest score of 1 is also in Condition 1 and
is ranked 2. Next comes a score of 2 also by a Condition 1 participant,
which is ranked 3. The two next lowest scores of 3 by one participant in
Condition 1 and also by one participant in Condition 2 are both assigned
the average of the next available ranks of 4 and 5 giving tied ranks of 4.5.
This procedure of assigning ranks is continued until all the scores have
been assigned overall rankings, regardless of which condition they come
from. For instance, a score of 4 by a participant in Condition 3 is ranked 6
while a higher score of 5 in Condition 2 is ranked 7. The top score of 11 by
a participant in Condition 3 is assigned the highest rank of 12 (note that
there are 12 participants, 4 in each group).
Table 12.2 Overall ranks of recall scores across conditions
Group 1 Group 2 Group 3
2 3 3 4.5 11 12
3 4.5 5 7 7 10
1 2 6 8.5 4 6
0 1 6 8.5 8 11
Totals of ranks for each condition
12.3
In order to decide whether the ranks are higher in one of the three con-
ditions, the totals of ranks are added up separately for each condition. The
aim is to check whether there are higher ranks in some conditions than
other conditions.
There are two stages in assessing scores from different participants.
The rst stage is to carry out an overall ranking of all individual scores,
regardless of which condition they are in. The second stage is to add the
totals for each of the three conditions separately. The totals of the ranks
for the three conditions are shown in Table 12.3.
?
(a) Would the rank totals in Table 12.3 support the hypothesis that groups Question 12.1
of participants will recall more very common words (Condition 3)
than rare words (Condition 1)?
(b) Calculate the means of the scores for each condition. Do these means
reect the same predicted differences between conditions?
The KruskalWallis test (unrelated) can deal with three or more con-
ditions. If there were four conditions the scores for the four conditions
would have to be assigned overall ranks for all scores, regardless of the
four conditions. The rank totals would then be added separately for each
of the four conditions.
Table 12.3 Totals of ranks for three conditions
Group 1 Group 2
(common words)
Group 3 (very
common words)
2 3 3 4.5 11 12
3 4.5 5 7 7 10
1 2 6 8.5 4 6
0 1 6 8.5 8 11
Totals 6 10.5 20 28.5 30 39
12.4
testing differences between three (or more) conditions..
Chart (inside back cover) as suggested, you will arrive at the Kruskal
Wallis test (unrelated). The KruskalWallis (unrelated) is the appropriate
non-parametric test for an unrelated design with three or more conditions.
Using the KruskalWallis test (unrelated)
12.5
When to use
The KruskalWallis test (unrelated) should be used for an unrelated design
when different participants are doing three (or more) conditions and the
data are ordinal.
Research hypothesis
The prediction was that the number of illustrations in a text would affect
participants recalls of ideas from the text.
Sample data
Three groups of different participants were presented with texts to read.
The rst group read texts which were highly illustrated (Condition 1), the
second group read tests with some illustrations (Condition 2), the third
group read tests with no illustrations (Condition 3). The dependent
variable was the number of ideas recalled from each text. The scores for
each group of different participants is shown in Table 12.4.
Notes on Table 12.4
1 Check the overall ranks assigned to the individual scores across all three
conditions.
2 The rank totals are added for each of the three conditions.
3 Note that in Table 12.4 there are unequal numbers of participants in the
three conditions (4 in Condition 1, 3 in Condition 2 and 4 in Condition
3). These unequal numbers conditions will affect how probabilities are
looked up in the statistical tables.
Rationale
The aim of the KruskalWallis test (unrelated) is to test whether the rank
totals are different for three groups of participants in three conditions.
If there are random differences between the rank totals for all three
conditions, as stated by the null hypothesis, we would expect the rank
totals to be approximately equal, because there would be equal numbers
of low and high ranks in each of the conditions. If the rank totals show a
preponderance of high or low ranks in one or other of the conditions, the
null hypothesis can be rejected.
Table 12.4 Number of ideas recalled for illustrated texts
Condition 1
(highly illustrated
text)
Condition 2
(text with some
illustrations)
Condition 3
(text with no
illustrations)
19 10 14 6 12 3.5
21 11 15 7 12 3.5
17 9 9 1 13 5
16 8 10 2
Rank totals 38 14 14
Means 18.25 12.67 11.75
Step-by-step instructions for calculating the KruskalWallis test
(unrelated)
These are given in Box F.
Box F
Step-by-step instructions for calculating the value of H
Rank all the scores across conditions 1
as an overall set of ranks, giving 1 to
the smallest score and so on.
See the Ranks columns in Table 12.4
which contain the overall ranks of all
scores taken together
Calculate rank totals for each condition. 2 T
1
= 38, T
2
= 14, T
3
= 14
Note the following symbols: 3
N = total number of participants N = 11
n = number of participants in each
group
n
1
= 4, n
2
= 3, n
3
= 4
T
2
= squares of the rank totals for
each condition
T
2
1
= 38
2
, T
2
2
= 14
2
, T
2
3
= 14
2
T
2
n
= sum of the squared rank totals
for each condition divided by
the number of subjects in that
condition
Calculate the value of H using the 4
formula
H =
12
N(N + 1)

T
2
n
3(N + 1) H =
12
11 12

38
2
4
+
14
2
3
+
14
2
4

3 12
=
12
132
1444
4
+
196
3
+
196
4

36
= 0.091(361 + 65.33 + 49) 36
= 43.2 36
= 7.2
Calculate the degrees of freedom: the 5
number of conditions (C) minus one.
df = C 1
= 3 1
= 2
Notes on calculations in Box F
1 The KruskalWallis test (unrelated) is a non-parametric test based
on ordinal data, in which overall scores are ranked and rank totals
calculated.
2 The calculations involve calculating the squares of the rank totals and
summing them to produce a sum of squares.
3 The formula in step 4 for calculating H has to take into account the
number of conditions and the number of participants in each group.
4 You are told to calculate degrees of freedom (df ). The df are the
number of conditions (C) minus 1 (in our example 3 1 = 2). Note that
you only need the df for experiments with relatively large numbers of
participants in each group.
Looking up the signicance of H in Tables F and E
Note that there are quite often special tables for small numbers of
conditions and participants and other tables for larger numbers.
Table F
Table F covers three conditions, with different numbers of participants in
each group, up to 5 participants in each group.
In Table F you have to look up the size of the groups in an experiment.
In our example the sizes of groups are n
1
= 4, n
2
= 3, n
3
= 4. When you look
up the group sizes in Table F it does not matter about the order of the
group sizes. The appropriate sizes are 4, 4, 3.
The calculated H has to be equal to or larger than the values in the
tables. The rst column for each set of group sizes gives the values of H.
The nal column gives the probabilities (p) associated with each value
of H. In our example the calculated H in Box F is 7.2. Look at the
probabilities for H for the 4, 4, 3 group sizes. There is an H of 7.1364,
which is the nearest to 7.2. The probability for this value of H is p < 0.01.
The null hypothesis can be rejected (p < 0.01)
Table E
In most experiments there will be more than ve participants in each
group. If you have more than three conditions and/or more than 5 partici-
pants per group, you will have to consult a more general statistical Table E
known as chi-square. To look up probabilities in Table E the df is the
number of conditions, C, minus 1. In our example df = 3 1 = 2. In Table E
look along the df = 2 row. The calculated H of 7.2 is larger than the value
of 5.99 for (p < 0.05). The null hypothesis can be rejected (p < 0.05).
Conclusions
It is essential to look up the means in Table 12.4 to check which texts with
illustrations resulted in more ideas being recalled. The means show that
more ideas were recalled from the highly illustrated texts in Condition 1
than from either of the other two texts. The KruskalWallis test
(unrelated) can only test whether there are overall differences between
conditions, which is the equivalent of a two-tailed hypothesis.
Note about names of non-parametric tests
12.6
Sometimes students nd it quite difcult to remember the names of
non-parametric tests. A tip about names for related and unrelated tests is
as follows, which is just a coincidence!
Related tests have just one name: Wilcoxon (related) for two conditions;
Friedman (related) for three conditions.
Unrelated tests were invented by two researchers so are named after
both: MannWhitney (unrelated) for two conditions; KruskalWallis
(unrelated) for three conditions.
Progress box twelve

In unrelated designs the scores in all three conditions are produced by
Individual scores are assigned overall ranks across all conditions.
The totals of ranks are added separately for each condition.
The appropriate statistical test for analysing unrelated ordinal data
for three conditions is the KruskalWallis test (unrelated).
Analysis of variance Part III
Introduction to ANOVA 13
Parametric tests
13.1
All the tests described in Part III are parametric tests. You will nd that
some of the concepts of analysing variance seem quite familiar.
The t tests, which were described in Chapters 8, 9 and 10, are also
parametric tests. Outlined below are some of the concepts that apply to all
parametric tests:
Numerical calculations on interval data. All parametric tests are based on
carrying out numerical calculations on interval data.
Ratios of variances. Parametric tests calculate ratios between predicted
variance and total variance.
Calculating variances. It is a general principle that variance is calculated
by squaring numbers and summing the squares in various combinations.
You will be given detailed information about which sums of squares to
calculate in the step-by-step instructions for each type of ANOVA.
Analysis of variance
13.2
The letters ANOVA stand for ANalysis Of VAriance. Analysing variance
is a basic feature of all parametric tests. One of the special advantages
of ANOVA is that total variance can be subdivided into other types of
variance. Total variance includes both predicted variance and variance due
to irrelevant variables. With ANOVA, rather than comparing predicted
variance against total variance, it is possible to subdivide total variance
into other kinds of variance.
Two subdivisions of total variance are: predicted or between-conditions
variance; and error variance due to unpredicted irrelevant variables.
Predicted variance is called between-conditions variance. This is because
Analysis of variance 84
the researcher has predicted total differences between conditions as a
result of the selected independent variable.
Variance due to unpredicted variables is called error variance. The
reason for this term is that, from the point of view of the experimenter,
anything that has not been predicted represents error.
Experimenters naturally try to reduce unpredicted error. One example
of unpredicted error might be if the same participants do all the experi-
mental conditions in the same order, leading to possible bias in par-
ticipants responses. A researcher routinely counterbalances the order
of conditions so as to remove order effects from possible error variance.
To summarize the argument so far, there are three different kinds of
variance in ANOVA. These are:
Predicted variance
Error variance
Total variance
It is important to note that predicted variance and error variance together
add up to total variance.
Sources of variance
13.3
Different kinds of variance are very commonly expressed as sources of
variance. This term is used to indicate that each kind of variance is
caused by a different type of variable. These variables are the sources of
the variance. The sources of variance in ANOVA are shown in Table 13.1.
In ANOVA the predicted between-conditions variance is calculated on
the basis of predicted differences in scores between conditions.
Total variance is calculated on the basis of differences in individual
scores produced by all participants in the experiment.
Error variance is calculated by subtracting predicted variance from total
variance. The rationale for this is that error variance is everything that is
left over after predicted variance has been accounted for.
Table 13.1 ANOVA sources of variance
Sources of variance Types of variance
Independent variables (predicted) Between-conditions variance
Irrelevant variables (unpredicted) Error variance
Combination of independent variables and
irrelevant variables
Total variance
In ANOVA, tests of signicance are based on the ratio of predicted,
between-conditions variance to error variance.
Degrees of freedom
13.4
There is one other important consideration that has to be taken into
account when calculating variances and when looking up statistics in
parametric statistics tables. This is the need to identify degrees of freedom
(df ).
You have been instructed to calculate df in some of the step-by-step
instructions for earlier tests. The rationale for degrees of freedom arises
from the notion that parametric tests calculate variances based on
variability in scores. It is essential that all scores are free to vary. The issue
is whether all scores from an experiment are equally variable.
The concept of degrees of freedom is quite a difcult one to grasp. An
experimental example might help.
Suppose you carry out an experiment and calculate the total of the
scores. When copying out the scores later, you forget to include one of the
scores from one of the six participants. You end up with the scores shown
in Table 13.2.
You do not need to panic because you know that, given ve scores and
the total of all six scores, you can calculate the forgotten score for par-
ticipant 6. All you have to do is to subtract the ve scores from the total of
75; the score for participant 6 could only have been 15. The implication is
that the score for that participant is predictable from knowing the other
scores and the total. So it has no freedom to vary.
Degrees of freedom take into account the fact that one of the scores
does not vary because it can be predicted from the other ve scores.
Degrees of freedom is a term to indicate all the other scores that can vary.
Table 13.2 Scores used as a basis
for calculating degrees of freedom
Participant Score
1 12
2 13
3 10
4 11
5 14
6
Total 75
Degrees of freedom are calculated as the total number of scores (N)
minus one score (N 1). In Table 13.2 there are six scores is but only ve
scores can vary. Subtracting 1 from 6 gives df = N 1 = 5. This conrms
that only ve scores are free to vary.
Subtracting 1 is a very general principle for calculating df. You will nd
that df are required for calculating ratios of variance. You will also nd
that df are important when looking up ratios in ANOVA statistical tables.
This is a very brief introduction to calculating degrees of freedom.
Sometimes df for combinations of variances have to be added or multi-
plied to obtain total df. In the step-by-step instruction for each type of
ANOVA you will be given detailed instructions about the calculation of
df.
?
In Table 13.3 there are four scores with one missing score and a total of 56. Question 13.1
(a) Calculate score 4.
(b) What are the degrees of freedom?
Requirements for parametric tests
13.5
There are three main requirements for parametric tests. These apply to
ANOVA and also to parametric t tests. These include (a) the measurement
of data, (b) the way in which the total scores are distributed and (c)
whether the scores are distributed equally in the experimental conditions.
We can deal quite easily with measurement of data. Because ANOVA
and t tests are parametric statistical tests, which involve numerical calcula-
tions based on sums of squares, the rst requirement is that that data must
be interval data, with equal intervals between scores on a continuous
numerical scale. It is assumed that remembering two items as opposed to
three items represents the same interval as that between remembering three
and four items.
Table 13.3 Calculating missing scores
Score 1 8
Score 2 12
Score3 20
Score 4
Total score 56
The question of the distribution of scores is more complex and we will
describe these concepts in the next section.
Normal distribution
13.6
A second requirement for parametric tests is that scores should be nor-
mally distributed.
The essence of a normal distribution of scores is that there tend to
be more middle-range scores than extreme scores. This is a very general
characteristic of distributions of scores. Think about the heights of the
adult population. You would expect to nd more people in the middle
range of heights (say 4.56 ft) than adults who are extremely tall (67 ft) or
extremely short (34 ft).
A normal distribution will be symmetrical, with equal numbers of
extreme scores at each end. Staying with the height example, the require-
ment is that there should be approximately equal numbers of rather short
people and very tall people.
One way to demonstrate this is to draw a histogram of all the scores
in an experiment, as shown in Figure 13.1. The possible scores are listed
along the bottom line. For every participant who scores a particular score
a box is added. The height of the boxes for each score represents the
number of participants who had that particular score. For example, six
participants had a score of 5.
?
(a) How many participants had a score of 1? Question 13.2
(b) How many participants had a score of 6?
What the curved line in Figure 13.1 shows is a theoretical normal distri-
bution for an unlimited number of scores.
Figure 13.1 Normal distribution
A third formal requirement for parametric data is termed homogeneity
of variance. The word homogeneity means sameness. This implies that
the variability of the scores in each condition should be approximately the
same.
The idea is that it would be difcult to make a comparison between
conditions based on very different distributions of scores. An example
would be if the scores in one condition were bunched around scores of 10,
11 and 12, compared with the scores in another condition ranging from
100 down to 1. Fortunately, such wide differences in variance between two
conditions are a rare occurrence.
It has been shown that, as long as there are equal numbers of partici-
pants in each condition, the results of parametric tests are fairly robust.
The word robust implies that, given that the data are approximately
normally distributed and there are not too wide discrepancies in condition
variances, the results of parametric tests will give a reasonably accurate
analysis of the data from an experiment. This is a good reason for
always allocating the same number of participants to each experimental
condition.
A rough check of whether scores from an experiment are approximately
normal is to draw a histogram of all the scores. You could draw histograms
for the scores in each condition to see whether distributions in each con-
dition are approximately equal. There are also more precise ways of ways
of checking a normal distribution and homogeneity of variance using
statistical software.
Computer statistical packages
13.7
Many students will have access to a commonly used computer statistical
package called SPSS. This statistical software is likely to be available in
psychology departments. It is also possible to download SPSS from the
Internet.
There are several advantages to using the SPSS computer package. It
carries out all the calculations needed to produce ratios of variance and
will even look up the statistical tables to provide a signicance level.
The SPSS package also automatically checks whether the data from an
experiment are approximately normally distributed and whether the scores
in each condition reach an acceptable homogeneity of variance.
Pressing a button on a computer to carry out all statistical calculations
and to deliver signicance levels may seem wonderful. However, note the
following words of warning.
1 You may not understand the rationale for selecting an appropriate stat-
istical test.
2 You may forget how to look up statistical tables for levels of
signicance.
3 Sometimes students calculate a level of signicance and do not check
the means to decide whether the signicant results support the pre-
dictions in the right direction as predicted by in the original research
hypothesis.
Despite all the advantages of SPSS, we will still be presenting boxes with
step-by-step instructions for calculating ANOVA. There are several
reasons for this decision. The rst reason is that you may not have access
to the SPSS computer program. An even more important reason is that it
is a good idea to get some idea of how sums of squares contribute to ratios
of variances.
ANOVA (unrelated) is the easiest to calculate so we will be presenting
this type of ANOVA rst. We will take you through the calculations at
every stage. This will help you to understand how the various formulae are
unpacked.
For later, more complex types of ANOVA, we will give the formulae and
the degrees of freedom necessary for calculating ratios and looking
them up in ANOVA statistical tables. You will be able to refer to the full
calculations in the ANOVA (unrelated) box if you need to calculate any of
the same sums of squares using a calculator or a computer package.
Progress box thirteen

The letters ANOVA stand for analysis of variance.
ANOVA analyses three sources of variance: predicted between-
conditions variance (predicted independent variable); error variance
(unpredicted irrelevant variables); and total variance (all the
variance in scores).
Ratios are calculated between predicted variance and error variance
by calculating sums of squares.
Degrees of freedom are calculated by deducting one predictable score
(N 1).
The main requirements for ANOVA parametric tests are that the
data are interval, the scores are normally distributed and that the
variances in conditions are approximately the same.
The SPSS computer program can be used to do calculations and
to check the requirements for parametric tests. But it is important to
understand the rationale for statistical tests.
One-way ANOVA (unrelated) 14
14.1
One-way ANOVA tests three (or more) conditions. In ANOVA (unrelated)
different groups of participants do each condition. Because different par-
ticipants do only one of the experimental conditions, it is not possible to
compare related scores for all conditions for each participant. Instead the
individual scores have to be treated as independent scores from different
participants.
The term one-way ANOVA indicates that there is only one independent
variable to be tested. This contrasts with later accounts of two-way
ANOVA, which are designed for two variables.
Denitions of variance
14.2
For ANOVA (unrelated) three sources of variance can be calculated.
These include:
Between-conditions variance (predicted)
Error variance (due to irrelevant variables and including individual
differences between participants)
Total variance (combined between conditions and error variance).
14.3
The answer for one-way ANOVA is one variable.
The answer is parametric because parametric statistical tests are suit-
able for analysing numerical interval data.
You will nd that if you answered the above ve questions in the Decision
Chart (inside back cover) as suggested, you will arrive at one-way ANOVA
(unrelated). ANOVA (unrelated) is the appropriate parametric test for a
one-way unrelated design with three (or more) conditions.
Using one-way ANOVA (unrelated)
14.4
When to use
One-way ANOVA (unrelated) should be used for an unrelated design when
different participants are doing three (or more) conditions testing one
independent variable and the data are interval.
Research hypothesis
The experimenter predicted that three presentation rates of lists of words
would have an effect on recall scores.
Sample data
Three different groups of six subjects were presented with lists of ten
words to learn. Group 1 were presented with words at a slow rate of
presentation of one word every 5 seconds (Condition 1), Group 2 with a
medium presentation rate of one word every 2 seconds (Condition 2)
and Group 3 with a fast presentation rate of one word every second
(Condition 3). The recall scores are shown in Table 14.1.
Notes on Table 14.1
1 The means have been calculated for each of the conditions.
2 The totals for the six participants in each condition have been
calculated.
3 The grand total (104) of the 18 scores in all three conditions has also
been calculated.
Rationale
The aim of ANOVA is to compare ratios of variances. The independent
variable of three presentation rates represents the predicted differences
between the three conditions (between-conditions variance). Differences in
scores due to irrelevant variables represent error variance.
The prediction is that the variance between conditions will be relatively
large compared with the error variance due to irrelevant variables. If
there are random differences between conditions, as stated by the null
hypothesis, the variance due to the predicted differences between con-
ditions will be small in relation to error variance. In this case, the null
hypothesis cannot be rejected.
Step-by-step instructions for calculating ANOVA (unrelated)
These are given in Box G.
Table 14.1 Number of words recalled for three presentation rates
Condition 1
(slow rate)
Condition 2
(medium rate)
Condition 3
(fast rate)
8 7 4
7 8 5
9 5 3
5 4 6
6 6 2
8 7 4
Totals (T) 43 37 24 104 Grand total
Means 7.17 6.17 4
Box G
Step-by-step instructions for calculating one-way F ratios
(unrelated)
Note the following symbols (see Table 14.1): 1
T
2
= sum of squared totals for each condition

T
2
= 43
2
+ 37
2
+ 24
2
n = number of participants in each condition n = 6
N = total number of scores N = 18
(
x)
2
= grand total squared (
x)
2
= 104
2
(
x)
2
N
= constant to be subtracted from all
sums of squares
104
2
18
=
10,816
18
= 600.89
x
2
= sum of each individual score squared
Calculate SS
bet
2
T
2
n

(
x)
2
N
=
43
2
+ 37
2
+ 24
2
6
600.89
= 632.33 600.89
= 31.44
Calculate SS
tot
3
x
2

(
x)
2
N
= 8
2
+ 7
2
+ 9
2
+ 5
2
+ 6
2
+ 8
2
+ 7
2
+ 8
2
+ 5
2
+ 4
2
+ 6
2
+ 7
2
+ 4
2
+ 5
2
+ 3
2
+ 6
2
+ 2
2
+ 4
2
600.89
= 664 600.89
= 63.11
Calculate SS
error
4
SS
tot
SS
bet
= 63.11 31.44
= 31.67
Calculate the degrees of freedom 5
df
bet
= number of conditions 1 df
bet
= 3 1 = 2
df
tot
= N 1 df
tot
= 18 1 = 17
df
error
= df
tot
df
bet
df
error
= 17 2 = 15
Divide each SS by df to obtain MS 6
MS
bet
=
SS
bet
df
bet
=
31.44
2
= 15.72
MS
error
=
SS
error
df
error
=
31.67
15
= 2.11
Notes on calculations in Box G
1 Note the denitions of the symbols and the calculation of a constant,
which will be subtracted from all sums of squares. This is the grand total
of all scores squared and divided by N.
2 Between-conditions variance (SS
bet
) is calculated by summing the
squared totals for each condition, dividing the sum by n (participants
in each condition) and subtracting the constant.
3 Total variance (SS
tot
) is calculated by summing the squares of all the
scores and subtracting the constant.
4 Error variance (SS
error
) is calculated by subtracting predicted variance
(SS
bet
) from total variance (SS
tot
).
5 Rules are given for calculating the degrees of freedom (df) by sub-
tracting 1 for each type of variance.
6 MS
bet
and MS
error
are calculated by dividing by the sums of squares (SS)
by the df.
7 The F ratio is calculated by dividing MS
bet
by MS
error
. Note that the df
are given for F (df
bet
= 2, df
error
= 15).
Looking up the signicance of F in Table G
Tables G(1) and G(2) are the appropriate tables for all ANOVA tests.
Table G(1) gives the values for p < 0.05. Table G(2) gives the values for p <
0.01. The calculated F has to be equal to or larger than these values. You
need to use the degrees of freedom for the between-conditions variance
Calculate the F ratio for MS
bet
over MS
error
7
allocating the correct df to MS
bet
and MS
error
F ratio for MS
bet
=
MS
bet
MS
error
F
2,15
=
15.72
2.11
= 7.45
The completed ANOVA table is shown in 8
Table 14.2.
Table 14.2 One-way ANOVA table (unrelated)
Sources of variance
Sums of
squares
Degrees of
freedom
Mean
squares
F
ratio
Presentation rate 31.44 2 15.72 F
2,15
= 7.45
Error 31.67 15 2.11
Total 63.11 17
and the error variance (see step 5 in Box G). (In our example these are
df
bet
= 2, df
error
= 15.)
In Table G(1) df
bet
is shown along the
1
row and df
error
is shown down
the
2
column. Locate 2 on
1
along the top line and follow the
2
column
down until you nd 15. Where they intersect is the critical value. In
Table G(1) the value for F
2,15
is 3.68 for p < 0.05. The calculated F of 7.45
in Box G is larger than 3.68, so the null hypothesis can be rejected (p <
0.05). The next step is to see whether the calculated value of F of 7.45 is
larger than the critical value of 6.36 in Table G (2) for df = 2, 15 (p < 0.01).
The calculated F is larger than the value of 6.36 so the null hypothesis can
be rejected (p < 0.01).
Conclusions
It is essential to look back at Table 14.1 to check the differences between
the means in the three conditions. There are differences between the
means for the three conditions with different presentation rates. This result
supports the research hypothesis that presentation rates of words lists
would have an effect on recall scores, with the fewest words recalled for the
fast presentation rate in Condition 3.
Note that ANOVA only tells you whether there are overall signicant
differences between the experimental conditions, which is equivalent to a
two-tailed hypothesis. In Chapter 16 tests for identifying the effects of
individual conditions will be discussed.
Progress box fourteen

In one-way ANOVA there is one independent variable.
In ANOVA (unrelated) the scores are produced by different
participants.
Three sources of variance are tested by ANOVA (unrelated):
Between-conditions variance error variance and total variance.
The appropriate parametric test for analysing three sources of
variance in an unrelated design is one-way ANOVA (unrelated).
One-way ANOVA (related) 15
Scores from same participant
15.1
One-way ANOVA tests three (or more) conditions. In ANOVA (related)
the same participants do all three conditions. Each participant produces
three scores, one in each of the three conditions. This makes it possible to
make a direct comparison between each participants three scores in all the
conditions.
The term one-way ANOVA indicates that there is only one independent
variable to be tested. This contrasts with later accounts of two-way
ANOVA, which are designed for two variables.
Denitions of variance
15.2
Because there are three related scores for each participant, it is possible to
calculate the variance between the three related scores for each participant.
Each participant will have a different set of three related scores. These
differences between participants provide the basis for calculating the
variance due to individual differences between participants.
In ANOVA (related) it is possible to identify an extra type of variance,
which represents individual variance between participants. One important
result of being to calculate individual variance between participants is that
for ANOVA (related) four kinds of variance can be calculated. These
include:
Between-conditions variance (predicted)
Individual variance (based on differences between related scores by each
participant)
Error variance (due to irrelevant variables)
Total variance (including all the above kinds of variances).
The net result of four types of variance is that the error variance has been
split into two. You will remember from Section 14.2 that error variance
was due to all irrelevant variables including individual differences between
participants. This was because in ANOVA (unrelated) there was no basis
for comparing related scores.
With ANOVA (related) the individual variance between participants
can be calculated as a separate source of variance. It is for this reason
that error variance no longer includes individual differences between
participants.
15.3
both conditions.
The answer is parametric because parametric statistical tests are suit-
able for analysing numerical interval data.
You will nd that if you answered the above ve questions in the Decision
Chart as suggested, you will arrive at one-way ANOVA (related). ANOVA
(related) is the appropriate parametric test for a one-way related design
with three (or more) conditions.
Using one-way ANOVA (related)
15.4
When to use
One-way ANOVA (related) should be used when the same participants are
doing three (or more) conditions testing one independent variable and the
data are interval.
Research hypothesis
The experimenter predicted that three presentation rates of lists of words
would have an effect on recall scores.
Sample data
The same six participants were presented with Condition 1 (words at a
slow rate of presentation of one word every 5 seconds), Condition 2
(words at a medium rate of one word every 2 seconds) and Condition 3
(words at a fast rate of one word every second). The recall scores for the
three conditions are shown in Table 15.1
?
What precautions should be taken by an experimenter to offset the Question 15.1
possible order effects of participants doing all three conditions in the same
order?
Notes on Table 15.1
1 The scores and totals in Table 15.1 are the same as in Table 14.1 in
Chapter 14.
2 Because all participants are doing all conditions the participants are
listed in the rst column. This makes it clear that each participant
produces three related scores, one in each condition.
3 The main difference with ANOVA (unrelated) is that total related
scores for the three conditions have been calculated for each of the six
participants. These totals appear in the last column of Table 15.1.
Participants
Condition 1
(slow rate)
Condition 2
(medium rate)
Condition 3
(fast rate)
Totals for
participants (T
s
)
1 8 7 4 19
2 7 8 5 20
3 9 5 3 17
4 5 4 6 15
5 6 6 2 14
6 8 7 4 19
Totals (T) 43 37 24 104 Grand total
Means 7.17 6.17 4
Rationale
The aim of ANOVA is to compare ratios of variances. The independent
variable of three presentation rates represents the predicted differences
between the three conditions (between-conditions variance). Differences in
scores due to irrelevant variables represent error variance. Differences
between participants represent individual variance.
The prediction is that the variance between conditions will be relatively
large compared with the error variance due to irrelevant variables. If
there are random differences between conditions, as stated by the null
hypothesis, the variance due to the predicted differences between con-
ditions will be small with to error variance. In this case, the null hypothesis
cannot be rejected. It is normal to compare individual variance with error
variance. But the experimenter will be pleased if there are no signicant
differences between individual participants.
Step-by-step instructions for calculating ANOVA (related)
These are given in Box H.
Box H
Step-by-step instruction for calculating one-way F ratios
(related)
Note the following symbols (see 1
Table 15.1):
T
2
= sum of squared total for
each condition
T
2
= 43
2
+ 37
2
+ 24
2
T
2
s
= sum of squared totals for
each participant
T
2
s
= 19
2
+ 20
2
+ 17
2
+ 15
2
+ 14
2
+ 19
2
n = number of participants n = 6
c = number of conditions c = 3
N = number of scores N = 18
(
x)
2
x)
2
= 104
2
(
x)
2
N
= constant to be subtracted
from all SS
104
2
18
=
10816
18
= 600.89
x
2
= sum of each individual
score squared
Calculate SS
bet
2
T
2
n

(x)
2
N
=
43
2
+ 37
2
+ 24
2
6
600.89
= 632.33 600.89
= 31.44
Calculate SS
subj
3
T
2
s
c

(
x)
2
N
=
19
2
+ 20
2
+ 17
2
+ 15
2
+ 14
2
+ 19
2
3
600.89
= 610.67 600.89
= 9.78
Calculate SS
tot
4
x
2

(x)
2
N
= 8
2
+ 7
2
+ 9
2
+ 5
2
+ 6
2
+ 8
2
+ 7
2
+ 8
2
+ 5
2
+ 4
2
+ 6
2
+ 7
2
+ 4
2
+ 5
2
+ 3
2
+ 6
2
+ 2
2
+ 4
2
600.89
= 664 600.89
= 63.11
Calculate SS
error
5
SS
tot
SS
bet
SS
subj
= 63.11 31.44 9.78
= 21.89
df
bet
= number of conditions 1 df
bet
= 3 1 = 2
df
subj
= number of subjects 1 df
subj
= 6 1 = 5
df
tot
= N 1 df
tot
= 18 1 = 17
df
error
= df
tot
df
bet
df
subj
df
error
= 17 2 5 = 10
MS
bet
=
SS
bet
df
bet
=
31.44
2
= 15.72
MS
subj
=
SS
subj
df
subj
=
9.78
5
= 1.956
MS
error
=
SS
error
df
error
=
21.89
10
= 2.189
Calculate F ratios for MS
bet
over 8
MS
error
and for MS
subj
over MS
error
allocating the correct df to the F ratios:
F ratios for MS
bet
=
MS
bet
MS
error
F
2,10
=
15.72
2.189
= 7.18
Notes on calculations in Box H
1 Some of the calculations of sums of squares are exactly the same as in
Box G in Chapter 14, as are the symbols in step 1, SS
bet
in step 2 and
SS
tot
in step 4. But there are also new sources of variance for ANOVA
(related).
2 In statistical tables it is universal to use the older notation SS
subj
to refer
to individual differences between participants (individual variance).
You will nd this notation used in all statistical textbooks (rather than
participant variance). To calculate SS
subj
the squared totals of T
s
are
summed and divided by the number of conditions, and the constant
subtracted
3 In step 5 all other variances are subtracted from SS
total
. Note that sub-
tracting SS
subj
as well results in a smaller SS
error
.
4 In step 6 the degrees of freedom are calculated for you. Because of the
extra SS
subj
the df are more complicated than for ANOVA (unrelated),
but you should follow the instructions.
5 In step 7 MS
bet
, MS
subj
and MS
error
are calculated by dividing the sums of
squares (SS) by the df.
6 In step 8 F ratios are calculated for MS
bet
and MS
subj
over MS
error
. Each
F ratio indicates the relevant df (2,10 and 5,10).
Tables G(1) and G(2) are the appropriate tables for all ANOVA tests. Table
G(1) gives the values for p < 0.05. Table G(2) gives the values for p < 0.01.
F ratios for MS
subj
=
MS
subj
MS
error
F
5,10
=
1.956
2.189
= 0.8935
The completed ANOVA is shown in 9
Table 15.2.
Table 15.2 One-way ANOVA (related)
Sources of variance
Sums of
squares
Degrees of
freedom
Mean
squares F ratios
Presentation rate 31.44 2 15.72 F
2,10
= 7.18
Subjects 9.78 5 1.956 F
5,10
= 0.8935
Error 21.89 10 2.189
Total 63.11 17
The calculated F has to be equal to or larger than the values in the tables.
You need to use the df for the between-conditions variance, the error
variance and the individual variance (see step 6 in Box H). (In our example
these are df
bet
= 2, df
error
= 10, df
subj
= 5.)
In Table G(1) df
bet
is shown along the 1 row on the top line and the
df
error
down the 2 column. Locate 2 on 1 along the top line and follow
down the 2 column to nd 10. Where they intersect is the critical value. In
Table G(1) the value for F
2,10
is 4.10. The calculated F of 7.18 in Box H is
larger than 4.10, so the null hypothesis can be rejected (p < 0.05). The next
step is to see whether the calculated F of 7.18 is larger than the value in
Table G(2) for df 2,10 (p < 0.01). 7.18 is smaller than the critical value of
7.56 so the null hypothesis cannot be rejected at this signicance level.
It is common practice to test the signicance of the F ratio for individual
variance, hoping that individual differences between participants will not
be a signicant source of variance. In our example, the calculated F for
SS
subj
of 0.8935 (df 5,10) is smaller than the value of 3.33 in Table G (1)
and so the differences between participants are not signicant.
Conclusions
It is essential to look back at the means in Table 15.1 to check whether
there are differences between the means for the three conditions. There
are differences between the means for the three conditions with different
presentation rates of words, which have been shown to be signicant
(p < 0.05). This supports the research hypothesis that presenting words at
faster or slower presentation rates would have an effect on recall scores.
Note that ANOVA only tells you whether there are overall signicant
differences between experimental conditions, which is the equivalent of
a two-tailed hypothesis. In Chapter 16 tests for identifying the effects of
individual conditions will be discussed.
Progress box fteen

In one-way ANOVA (related) there is one independent variable.
In ANOVA (related) the scores are produced by the same
participants.
Four sources of variance are tested by ANOVA (related): between-
conditions variance, individual variance, error variance and total
variance.
The appropriate parametric test for analysing four sources of
variance in a related design is one-way ANOVA (related).
Comparisons between ANOVA
16
conditions
Overall differences in ANOVA
16.1
In the conclusion sections in Chapters 14 and 15, we made the point that
ANOVA only provides information about overall differences between
three experimental conditions. However clear the differences seem to be
between the means for individual conditions, ANOVA tables can only test
whether there are overall signicant differences. ANOVA cannot identify
which individual conditions contributed most to these overall differences.
What researchers would like to do is identify which individual con-
ditions contribute to the signicant overall differences in scores in the
experimental conditions. In order to achieve this they need to identify a
pattern of the mean scores for all the conditions.
The rst step is to look at the pattern of mean scores to see if some of
them indicate that not all the experimental conditions have contributed to
the overall ANOVA in exactly the same way.
Graphs showing patterns of mean scores
16.2
One method for displaying the pattern of mean scores is to draw them in
the form of a graph. A graph of mean scores is shown in Figure 16.1.
Every graph has a horizontal line and a vertical line; each of these is called
an axis.
In the case of Figure 16.1 the three mean scores are for three rates of
presenting lists of words: A (fast), B (medium) and C (slow). The three
presentation rates are listed along the horizontal axis and the mean scores
are listed up the vertical axis. For each presentation rate the mean score is
marked.
Looking at these mean scores it is obvious that, rather than the three
conditions contributing equally to the one-way ANOVA, there is a denite
trend between the mean scores, from lowest to highest. The fast rate has
the lowest mean score, the slow rate the highest mean score, with the
medium rate having a mean score in between.
In the examples of one-way ANOVA given in Chapters 14 and 15 the
aim was to test the effects of different rates of presentation. The point
about ANOVA is that it can only test overall differences between the
different presentation rates. The analysis in this chapter makes it possible
to investigate the effects of individual experimental conditions on mean
scores.
Comparisons between individual conditions
16.3
Let us summarize the argument so far. The rst step is to check that there
are overall signicant differences between the experimental conditions, as
revealed by ANOVA tables.
If there are signicant overall differences between conditions, a
researcher may want to carry out further investigations in order to test
comparisons between the individual conditions. The aim is to test whether
the scores for individual conditions support the predicted differences in
scores.
Because the comparisons between individual conditions are carried
out after the main ANOVA analysis, these comparisons are known as post
Figure 16.1 Graph showing recall scores for three presentation rates
Comparisons between ANOVA conditions 105
hoc comparisons. Post hoc is a Latin phrase meaning after the event,
which tells us that the comparisons are after the original ANOVA.
These comparisons are also called unplanned comparisons. This is
because it is only after the means of the individual conditions have been
checked that a researcher may decide that it may be worth investigating
differences between individual conditions.
Multiple comparisons between conditions
16.4
In order to explore any differences between individual conditions, the
researcher needs to test comparisons between pairs of individual con-
ditions. These are often referred to as multiple comparisons because several
comparisons can be made between pairs of individual conditions. They
are sometimes also referred to as contrasts because they contrast the effects
of pairs of individual conditions.
There are three possible pairs of contrasts between three individual
conditions:
Contrast between Condition 1 and Condition 2
One relevant contrast is to test whether the recall scores for Condition 1
(slow rate) are higher than the recall scores for Condition 2 (medium
rate). The second relevant contrast is to test whether the recall scores for
Condition 2 (medium rate) are higher than the recall scores for Condition
3 (fast rate).
These two contrasts represent the predicted differences in recall scores
from highest in Condition 1 (slow) to fewest in Condition 3 (fast). The
mean recall scores in the graph in Figure 16.1 reect the predicted scores,
with the highest mean recall scores for C (slow rate).
Using t tests for pairs of conditions
16.5
The next question that has to be considered is how to go about testing for
contrasts between individual conditions. The basic method is to use t tests
between each pair of conditions.
You will remember that t tests are designed for testing differences
between two conditions. This is why t tests are ideal for testing com-
parisons between pairs of individual conditions.
The ANOVA for this experiment was unrelated because groups of
different participants were doing each of the three conditions. This
means that the appropriate t test for pairs of conditions is the t test
(unrelated).
One t test (unrelated) can be used to test the rst contrast that scores for
slow rates are different from the scores for medium rates. Another t test
(unrelated) can be used to test the second contrast that scores for medium
rates are different from the scores for fast rates.
Note that, although ANOVA can only test overall differences, the
researcher has made specic predictions about the predicted contrasts
between individual conditions. For this reason, these contrasts can be
treated as one-tailed hypotheses when looking up the calculated t tests
(unrelated) in the appropriate statistical table.
Adjusting signicance levels
16.6
The need to adjust signicance levels arises because several different tests
are being carried out on the same data. There is the original ANOVA plus
the possibility of three t tests between pairs of scores on individual
conditions.
Statistical tests use the conventional level of signicance (p < 0.05). This
represents 1 in 20 probability of random variability. But if a researcher
keeps testing the same data, there may be 1 in 20 probability that one of
these tests will only appear to be signicant. It may be the case that one of
the contrasts is accepted as signicant even though the scores are random,
as stated by the null hypothesis.
To understand what this means we strongly recommend you to look
back to Section 4.4, which introduced the notion of one-tailed and two-
tailed hypotheses. The point was made that a one-tailed hypothesis pre-
dicts differences in scores in conditions in one direction only. A one-tailed
hypothesis may have a specic probability of random variance (p < 0.05).
A two-tailed hypothesis makes a prediction that the differences in scores
in conditions might go in either direction, in favour of either one condition
or the other. A two-tailed hypothesis has double the probability of random
variance because two hypotheses are being tested, one in one direction and
one in the other direction. When a two-tailed hypothesis is being tested,
the signicance levels have to be adjusted to allow for two hypotheses
being tested.
The need for stricter levels of signicance for two hypotheses in a
two-tailed test applies even more strongly when there are several multiple
comparisons between possible pairs of conditions.
It is essential to understand that, if the same signicance levels were
used for testing all pairs of contrasts between conditions, we would be
biasing the analysis in favour of having several goes at obtaining signi-
cance based on the same data.
Just as with the two-tailed hypothesis, signicance levels have to be
stricter, depending on the number of comparisons that are being tested.
The t tests for comparisons that we will be describing for testing several
contrasts on the same data are based on the general principle that signi-
cance levels have to be adjusted for multiple comparisons. The researcher
may be interested in testing only one or two of the contrasts between
conditions. But from a statistical point of view, other possible contrasts
between conditions also have to be taken into account.
Bonferroni test
16.7
The Bonferroni test uses a very direct way of adjusting signicance levels
in relation to the number of comparisons that could be carried out on
three possible contrasts.
The conventional signicance level is p < 0.05. The Bonferroni approach
to obtaining stricter levels of signicance is achieved by dividing the
0.05 signicance level by the number of possible comparisons. For three
contrasts, 0.05 has to be divided by 3, giving 0.017. This represents an
adjusted signicance level (p < 0.01), which applies to all the t tests for
testing the contrasts between individual conditions. It is necessary to
achieve the adjusted signicance level of p < .0.01 for the t tests to be
counted as being signicant at the conventional p < 0.05.
As an example, suppose the calculated t test (unrelated) used for testing
the contrast between Condition 1 (slow) and Condition 2 (medium) in
Table 16.2 turned out to be t = 2.65 (df = 18). We need to test this value of
t at the adjusted signicance level (p < 0.01). If the adjusted signicance
level of (p < 0.01) is not achieved, the null hypothesis cannot be rejected at
the signicance level of p < 0.05.
In Table C for t tests, when you look along the row for df =18, the
calculated t of 2.65 is larger than the adjusted value of 2.552 for one-tail
signicance (p < 0.01). Remember that the researcher can only claim a
signicant contrast between the recall scores in Condition 1 (slow) and
Condition 2 (medium) at the conventional p < 0.05.
Tukeys honestly signicant difference
16.8
When to use
Tukeys honestly signicant difference (HSD) or Q test takes into account
all possible contrasts between pairs of conditions. Even if a researcher
only wants to test one contrast, the HSD is based on all possible contrasts.
Research hypothesis
The prediction is that there will be differences between individual presenta-
tion rates.
Sample data
Table 16.1 gives the overall differences in scores that were analysed in
Chapter 14.
These data were tested by a one-way ANOVA (unrelated) because
different participants did the three conditions. The overall differences
between the three conditions were found to be signicant (p < 0.01). The
only contrast worth investigating in Table 16.1 is the comparison between
Condition 2 (medium rate and Condition 3 (fast rate).
Rationale
The Tukey HSD is used to test differences between pairs of contrasts
between individual conditions. The basic method that HSD uses for
adjusting for several comparisons on the same data is to take into account
the number of conditions that could have been made. The prediction is
that there will be relatively large differences between conditions in relation
to error variance. If there are random differences between the pairs of
individual conditions, as stated by the null hypothesis, the null hypothesis
cannot be rejected.
Condition 1 (slow
rate)
Condition 2
(medium rate)
Condition 3
(fast rate)
8 7 4
7 8 5
9 5 3
5 4 6
6 6 2
8 7 4
Total (T) 43 37 24
Means 7.17 6.17 4
Step-by-step instructions for calculating Tukeys Q
These are given in Box I.
Looking up the signicance of Q
You can see that the calculation of Q is quite similar to the t test
(unrelated) comparing the difference in means against an estimate of total
error variance. The difference is in the tables used to test the calculated
values of Q. The df are the same as those for the error variance in Box G
in Chapter 14 (df = 15).
Box I
Step-by-step instructions for calculating Tukeys Q
Note the following symbols 1 (see Table 16.1)
M
1
= mean for medium rate M
1
= 6.17
M
2
= mean for fast rate M
2
= 4
(Note that the lower mean has to be
subtracted from the higher mean)
n = number of participants in each condition n = 6
The formula for Q is: 2
Q =
M
1
M
2
MS
error
n
M
1
M
2
= 6.17 4 = 2.17
MS
error
can be found in the can be found in the
original ANOVA (unrelated) in Box G in
Chapter 14).
MS
error
= 2.11
Q =
6.17 4
2.11
6
=
2.17
0.35
=
2.17
0.59
= 3.67
The tables for Q can be found in several textbooks. The Q table is closely
based on the t table. The difference is that there is also a column for k, the
number of possible contrasts. This means that the number of possible
contrasts can be taken into account when calculating signicance levels.
In our example k = 3 for the three possible contrasts between the three
conditions for presentation rates. Because of the stricter requirement for
achieving signicance, the contrast between Condition 2 (medium rate)
and Condition 3 (fast rate) was only just signicant (p < 0.05).
Conclusions
The contrast between Condition 2 (medium rate) and Condition 3 (fast
rate) is signicant (p < 0.05), taking into account the three possible con-
trasts that could have been made. The very small difference between the
means of Condition 1 and Condition 2 are obviously not signicant.
Note about comparisons between conditions in ANOVA (related)
The instructions given in Box I refer to unrelated ANOVA. For ANOVA
(related) the principle of taking into account the contrasts that could have
been made is the same. Because the ANOVA is related, the comparisons
between individual conditions involve carrying out t tests (related),
adjusting the signicance level depending on the number of contrasts that
could have been made.
Introduction to two-way ANOVA 17
Comparison with one-way ANOVA
17.1
In Chapters 14 and 15 you were introduced to one-way ANOVA
(unrelated) and one-way ANOVA (related). It was explained that the term
one-way ANOVA indicates that there is only one independent variable.
There can be several conditions on one independent variable. An
example would be an experiment in which there is one independent
variable of three, or even four, different types of word lists.
In this chapter we discuss the possibility of investigating two independ-
ent variables. Instead of the experimenter selecting just one independent
variable of lists of words, the experimenter will also be able to select a
second variable.
Let us take as an example an experiment that tests the effect of a reading
scheme on childrens reading scores. In one-way ANOVA it would only be
possible to investigate one independent variable, either giving the children
a reading scheme or not giving them a reading scheme.
In two-way ANOVA two variables can be tested. A second variable
could be whether the children are good readers or less good readers.
Two-way ANOVA
17.2
Using two-way ANOVA it is possible to investigate the effects of the
reading scheme, the performance of good and less good readers and the
crucial question of whether good and less good readers are affected by
the reading scheme in different ways.
Two-way ANOVA can test the possibility that there might be a dif-
ferential effect of one independent variable on another. This is one of
the great advantages of two-way ANOVA. These differential effects are
known as interactions between the two variables. Interactions will be fully
discussed in the justication for two-way ANOVA experimental
designs.
Combining the two variables results in four conditions, as shown in
Table 17.1. Condition 1 shows the combination of less good readers with
no reading scheme; Condition 2, good readers with no reading scheme;
Condition 3, less good readers with a reading scheme; and Condition 4,
good readers with a reading scheme.
Two-by-two tables
17.3
In order to demonstrate that two-way ANOVA is testing two independent
variables, this is usually expressed as a two-by-two table. These are also
called 2 2 tables.
In 2 2 tables each of the independent variables is shown separately so
that the effects of the two variables can be identied. The 2 2 table in
Table 17.2 is another way of displaying the experimental conditions given
in Table 17.1.
Table 17.2 shows the effects of two independent variables, each of which
has two conditions. The rst independent variable is the reading scheme
(Variable A: reading scheme or no reading scheme). The second independ-
ent variable is reading skills (Variable B: good or less good readers).
The top left-hand box in Table 17.2 represents Condition 1 in Table
17.1: less good readers who are not given a reading scheme. The top right-
hand box represents Condition 3: less good readers with a reading scheme.
Boxes in 2 2 tables are usually labelled as cells so that there is no need
to refer to the top left-hand box. The boxes in Table 17.2 are labelled as
cells 1 to 4, as shown in Table 17.3.
Table 17.1 Four conditions in a two-way ANOVA
Condition 1 Condition 2 Condition 3 Condition 4
Less good readers Good readers Less good readers Good readers
No reading scheme No reading scheme Reading scheme Reading scheme
Table 17.2 2 2 table for four conditions
Variable B: Reading skills Variable A: Reading scheme
No reading scheme Reading scheme
Less good readers Condition 1 Condition 3
Good readers Condition 2 Condition 4
The next step is to put some numbers into the 2 2 table. Let us suppose
that we had carried out an experiment and had recorded scores for each
of the four conditions. These scores have been entered into the appropriate
cells in Table 17.4.
In Table 17.4 the totals for the Variable A reading scheme conditions
and the totals for the Variable B good versus less good readers conditions
have also been calculated.
The total for both conditions in which there is no reading scheme adds
up the scores for both the less good readers (2) and the good readers (8):
2 + 8 = 10. The equivalent total for conditions in which there is a reading
scheme is the total of less good readers (7) and good readers (8), adding up
to 15.
The difference between the totals for no reading scheme (10) and
reading scheme (15) shows the overall difference between these two
conditions regardless of whether the children were good or less good
readers. This overall difference is known as a main effect because it com-
pares the overall scores for Variable A (no reading scheme versus reading
scheme).
?
Which main effect is represented by the totals of 9 and 16 at the right-hand Question 17.1
side of Table 17.4?
Table 17.3 Labelled cells in 2 2 table
No reading scheme Reading scheme
Less good readers Cell 1 Cell 3
Good readers Cell 2 Cell 4
Table 17.4 2 2 table of score
Variable B Variable A
No reading scheme Reading scheme Totals
Less good readers 2 7 9
Good readers 8 8 16
Totals 10 15
Interpreting interactions
17.4
It is obviously interesting to look at the main effects in a 2 2 experiment
as shown in Table 17.4. These represent the differences between the effects
of Variable A, no reading scheme versus reading scheme (total scores of 10
versus 15, showing the main effect of higher scores for the reading
scheme). Table 17.4 also shows the differences between Variable B, less
good and good readers (9 versus 16, showing the main effect of good
readers performing better). However, it could be argued that these main
effects could just as well have been demonstrated in two separate
experiments. One experiment could have looked at performance with and
without the reading scheme. Another experiment could have looked at the
performance of less good and good readers.
The whole point of two-way ANOVA is that two independent variables
can be investigated simultaneously to see whether the two variables
inuence each other. What we want to know is whether it is the less
good readers or the good readers who benet most from the reading
scheme.
The experimental hypothesis was that less good readers would benet
more from the reading scheme. Good readers are already so good at
reading that they would not benet so much from a reading scheme.
If you look at Table 17.4, it is clear that the overall performance of the
good readers is better than the less good readers overall (scores of 16
versus 9). Now look at the individual cells in the table. It is easy to see that
the scores for the good readers are no different regardless of whether they
had a reading scheme or not (8 versus 8). On the other hand, the less good
readers, when given a reading scheme, improved from 2 to 7.
This could be interpreted as showing that good readers are unlikely to
show an improvement due to a reading scheme while less good readers are
likely to improve their reading performance when they are given a reading
scheme. Another way of expressing this is that scores on Variable A (no
reading scheme versus reading scheme) are affected by Variable B (good or
less good readers).
It is only the less good readers whose performance is affected by the
reading scheme. The fact that performance on one variable is affected
by performance on the other variable demonstrates that there is an
interaction between the two variables.
Interactions are often thought to be a very difcult concept to grasp.
But all they mean is that two independent variables interact in the sense
that one independent variable is affected by another independent variable.
An alternative way of putting this is that participants behave differently on
one of the variables as a direct result of the second variable.
Using graphs to demonstrate interactions
17.5
One good way of indicating whether there is an interaction between
variables is to draw a graph of the data. In Figure 17.1 the vertical axis is
labelled Reading scores. The horizontal axis is labelled No reading scheme
and Reading scheme.
The aim is to plot the reading scores of the less good and good readers
against the no reading scheme and reading scheme conditions. We use
the total scores in Table 17.4. For good readers the reading scores were 8
for both the no reading scheme and reading scheme conditions. Scores for
good readers are both plotted against the score of 8 on the reading scores
vertical axis.
For less good readers the reading scores were 2 for no reading scheme
and 7 for the reading scheme, so for less good readers their reading scores
are plotted at 2 and 7 on the vertical axis.
In order to demonstrate whether there is an interaction the normal
practice is to draw lines connecting the reading scores for good and less
good readers. The line connecting the reading scores for the good readers
is at because good readers got exactly the same score whether they were
given a reading scheme or not.
In contrast, the line connecting the scores for less good readers shows
a denite slope, rising from 2 for the no reading scheme condition to 7 for
Figure 17.1 Graph showing interaction
the reading scheme condition. The rising line demonstrates an improve-
ment in reading scores for the less good readers. These lines indicate that
there is quite likely to be an interaction between the amount of improve-
ment of good readers and less good readers.
We will now give another example where there is no interaction. Dif-
ferent scores for the four conditions are shown in Table 17.5. The scores
in Table 17.5 are plotted on the graph in Figure 17.2.
In the graph in Figure 17.2 good readers have higher reading scores than
less good readers (13 versus 9). But in this case both good readers and less
good readers benet from the reading scheme.
The less good readers improve from 2 to 7 and the good readers improve
from 4 to 9. These equal rates of improvements are shown in Figure 17.2
as the same slopes for both good and less good readers.
Figure 17.2 Graph showing no interaction
Table 17.5 2 2 table of scores
No reading scheme Reading scheme Totals
Less good readers 2 7 9
Good readers 4 9 13
Totals 6 16
These equal slopes indicate that there is no hint of any interaction
between the two variables. Both groups of readers have shown the same
amount of improvement from the reading scheme.
You will probably have realized by now that a graph is simply another
way of representing a 2 2 table. The graph in Figure 17.1 is equivalent
to the 2 2 table of scores in Table 17.4. The graph in Figure 17.2 is
equivalent to the 2 2 table of scores in Table 17.5.
Progress box seventeen

In one-way ANOVA only one independent variable (IV) can be
tested.
In two-way ANOVA two IVs can be tested simultaneously.
Main effects can be identied for each IV separately.
Interactions occur when different values of one IV affect per-
formance on a second IV.
Graphs are a good method for demonstrating interactions.
Two-way ANOVA (unrelated) 18
18.1
In two-way ANOVA (unrelated) different participants do each condition.
Because different participants do only one of the experimental conditions,
it is not possible to calculate variance by comparing related scores for each
participant.
In two-way ANOVA two independent variables (IVs) are tested. In
two-way ANOVA (unrelated) two groups of different participants do the
two conditions on one independent variable, while Two groups of different
participants do the two conditions on the other independent variable.
Thus there are four groups in all.
An example would be an experiment in which the rst IV is a com-
parison between learning lists of long or short words. The second IV is
a comparison between fast and slow rates of presenting the lists of words,
as shown below.
Variable A (length of words)
Condition A
1
(short words)
Condition A
2
(long words)
Variable B (presentation rates)
Condition B
1
(fast rate)
Condition B
2
(slow rate)
Two groups of different participants will be doing the two Variable A
conditions (length of words). Two other groups of different participants
will be doing the two Variable B conditions (presentation rates).
This means that there will be four groups of different participants
each doing one of the four combinations listed in Table 18.1. We have
indicated this by showing the four different groups doing one of the four
conditions.

?
Suppose Variable A is a comparison between illustrated texts (A
1
) and Question 18.1
non-illustrated texts (A
2
). Variable B is a comparison between texts with a
title (B
1
) and texts with no title (B
2
).
Following the example in Table 18.1, ll in the combinations of the two
variables for four groups of different participants.

Denitions of variances
18.2
For two-way ANOVA (unrelated), where there are two independent
variables, there are ve sources of variance that can be calculated:
Between conditions on Variable A (predicted)
Between conditions on Variable B (predicted)
Interaction between A and B (predicted), which are denote by A B
Error variance (due to irrelevant variables, including individual
differences between participants)
Total variance (including all the above sources of variance).
18.3
Table 18.1 Four groups of different participants doing four conditions
Group 1 Short words (A
1
) presented at a fast rate (B
1
)
Group 2 Short words (A
1
) presented at a slow rate (B
2
)
Group 3 Long words (A
2
) presented at a fast rate B
1
)
Group 4 Long words (A
2
2
)
Group 1
Group 2
Group 3
Group 4
You will nd that if you answered the above three questions in the Deci-
sion Chart as suggested, you will arrive at two-way ANOVA (unrelated).
This is the appropriate parametric test for a two-way unrelated design
testing two independent variables.
Using two-way ANOVA (unrelated)
18.4
When to use
Two-way ANOVA (unrelated) should be used for an unrelated design when
different participants are doing two conditions on two independent vari-
ables and the data are interval.
Research hypothesis
It was predicted that there would be an interaction between Variable A
and Variable B. More short words would be recalled at a fast rate of
presentation and more long words at a slow rate of presentation.
Sample data
Sixteen participants are selected. Four groups of different participants
are allocated to the four combinations of Variable A conditions and
Variable B conditions. The scores and means are shown in Table 18.2.
Table 18.2 Recall score for four participants on each condition
A
1
(short words) A
2
(long words)
B
1
(fast rate) B
2
(slow rate) B
1
(fast rate) B
2
(slow rate)
9 4 5 7
8 3 3 5
6 3 3 6
7 5 4 7
Means 7.5 3.75 3.75 6.25
The scores in Table 18.2 are entered into four cells as shown in
Table 18.3.
Notes on Table 18.3
1 Four scores are shown for the four participants in each cell of the
table. There are four participants in cell A
1
/B
1
; four different partici-
pants in cell A
1
/B
2
; four different participants in cell A
2
/B
1
; four different
participants in cell A
2
/B
2
.
2 The totals have been added for the four scores in each cell.
3 The total scores for the two Variable A conditions have been added
(along the bottom line) and the total scores for the two Variable B
conditions (at the right-hand side).
4 These totals have been added to produce a grand total of 85.

?
(a) Check that the scores for each participant in the four cells in Table Question 18.2
18.3 are the same scores as those shown for the combinations of
Variable A and Variable B conditions for the four groups shown in
Table 18.2.
(b) Check in Table 18.3 that the totals of scores in the four cells and
the totals for the Variable A and Variable B conditions all add up to
the grand total of 85.

Table 18.3 Recall scores for two variables (unrelated)
Variable B (rate
Variable A (word length)
of presentation) A
1
(short words) A
2
(long words) Totals B
B
1
(fast rate) 9 8 6 7 5 3 3 4
30 15 45
B
2
(slow rate) 4 3 3 5 7 5 6 7
15 25 40
Grand
total
Totals A 45 40 85
Rationale
The aim of ANOVA is to compare ratios of variance. The between-
conditions variance for Variable A, Variable B and the interaction A B
represent predicted variances between conditions. Differences in scores
due to irrelevant variables represent error variance. The prediction is that
there will be relatively large differences between conditions compared with
error variance. If there are random differences between conditions, as
stated by the null hypothesis, the predicted differences between conditions
will be small in relation to error variance. In this case, the null hypothesis
cannot be rejected.
Step-by-step instructions for calculating two-way ANOVA
(unrelated)
These are given in Box J.
Box J
Step-by-step instructions for calculating 2 2 F ratios
(unrelated)
Note the following symbols (see Table 15.5): 1
T
2
a
= sum of A squared totals

T
2
a
= 45
2
+ 40
2
T
2
b
= sum of B squared totals

T
2
b
= 45
2
+ 40
2
T
2
ab
= the sum of AB squared totals

T
2
ab
= 30
2
+ 15
2
+ 15
2
+ 25
2
n = number of subjects in each condition n = 4
a = number of conditions for variable A a = 2
b = number of conditions for variable B b = 2
(
x)
2
x)
2
= 85
2
(
x)
2
N
= constant to be subtracted from all SS
85
2
16
= 451.5625
x
2
= sum of each individual score squared
Calculate SS
A
2
T
2
a
nb

(
x)
2
N
=
45
2
+ 40
2
4 2
451.5625
= 453.125 451.5625
= 1.5625
Calculate SS
B
3
T
2
b
na

(
x)
2
N
=
45
2
+ 40
2
4 2
451.5625
= 453.125 451.5625
= 1.5625
Calculate SS
AB
4
T
2
ab
n

(
x)
2
N
SS
A
SS
B
=
30
2
+ 15
2
+ 15
2
+ 25
2
4
451.5625 1.5625
1.5625
Note that you have already calculated SS
A
and
SS
B
and the constant.
= 493.75 451.5625
1.5625 1.5625
= 39.0625
Calculate SS
tot
5
x
2

(
x)
2
N
= 9
2
+ 8
2
+ 6
2
+ 7
2
plus all
other individual scores
squared 451.5625
= 507 451.5625
= 55.4375
Calculate SS
error
6
SS
tot
SS
A
SS
B
SS
AB
= 55.4375 1.5625
1.5625 39.0625
13.25
df
A
= number of A conditions minus 1 df
A
= 2 1 = 1
df
B
= number of B conditions minus 1 df
B
= 2 1 = 1
df
AB
= df
A
df
B
df
AB
= 1 1 = 1
df
tot
= N 1 df
tot
= 16 1 = 15
df
error
= df
tot
df
A
df
B
df
AB
df
error
= 15 1 1 1 = 12
MS
A
=
SS
A
df
A
MS
A
= 1.5625
MS
B
=
SS
B
df
B
MS
B
= 1.5625
MS
AB
=
SS
AB
df
AB
MS
AB
= 39.0625
MS
error
=
SS
error
df
error
MS
error
= 1.104
Calculate F ratios for MS
A
, MS
B
and MS
AB
9
allocating the correct df to the F ratios
F ratio for MS
A
=
MS
A
MS
error
F
1,12
= 1.415
F ratio for MS
B
=
MS
B
MS
error
F
1,12
= 1.415
F ratio for MS
AB
=
MS
AB
MS
error
F
1,12
= 35.38
The completed ANOVA table is shown 10
in Table 18.4
Notes on calculations in Box J
1 Note the denitions of the symbols and the calculation of the constant,
which is to be subtracted from all sums of squares.
2 SS
A
is based on the two Variable A totals shown in Table 18.3.
3 SS
B
is based on the two Variable B totals shown in Table 18.3.
4 SS
AB
is based on the totals of the four scores in each cell in Table 18.2
minus SS
A
and SS
B
and, of course, the constant.
5 Total variance (SS
tot
) is calculated from all the individual scores.
6 SS
error
is calculated by subtracting all the other variances from SS
tot
.
7 Calculate the df as instructed. The only point to notice is that df
AB
is the
result of multiplying df
A
by df
B
.
8 MS values are calculated by dividing SS by the relevant df.
9 F ratios are calculated by dividing the three predicted MS variances of
A, B and A B by MS error.
Tables G(1) and G(2) are the appropriate tables for all ANOVA tests.
Table G(1) gives the critical vales for p < 0.05. Table G(2) gives the critical
Table 18.4 Two-way ANOVA table (unrelated)
Sources of variance
Sums of
squares
Degrees of
freedom
Mean
squares F ratios
Variable A (word length) 1.5625 1 1.5625 F
1,12
= 1.415
Variable B (presentation rate) 1.5625 1 1.5625 F
1,12
= 1.415
A B (interaction) 39.0625 1 39.0625 F
1,12
= 35.38
Error 13.25 12 1.104
Total 55.4375 15
values for p < 0.01. The calculated F has to be equal to or larger than
the critical values. You need to use the degrees of freedom for all three
between-conditions variances A, B, AB, and the error variance. You
can use the df from step 7 in Box J (in our example df
A
= 1, df
B
= 1, df
AB
1,
df
error
= 12). What makes this easier is that all the variances have the same
df. So we only have to look up the values for df = 1,12.
In Table G (1) df for variables A,B,AB, is shown along the
1
row and
df error is shown down the
2
column. Locate 1 on
1
along the top line
and follow the
2
column down until you nd 12. Where they intersect is
the critical value for F
1,12
of 4.75 (p < 0.05). The calculated F for A and B
of 1.415 is obviously much smaller than 4.75. But the calculated F for the
A B interaction of 35.38 is larger than the critical value of 4.75. The
F of 35.38 is also larger than the critical value in Table G(2) for df = 1,12
of 9.33. So the null hypothesis for the A B interaction can be rejected
(p < 0.01).
Conclusions
The A B interaction between Variable A and Variable B is signicant
(p < 0.01). If we look at the means in Table 18.2, it is clear that there are
higher means for short words at a fast rate (7.5) and for long words at a
slow rate (6.25). One possible reason for this signicant A B interaction
is that long words take longer to memorize and so benet from a slow rate
of presentation. Short words take less time and so with a slow rate they
may be forgotten. Short words are likely to benet from a faster rate of
presentation.
Using graphs
18.5
Interactions are often demonstrated by graphs. The output of an SPSS
computer program is an ANOVA table (see Table 18.4) and usually a graph
as well. But it is useful to plot a graph for yourself in order to understand
the signicance of an interaction.
In the graph in Figure 18.1 the fast rate and slow rate are shown along
the bottom line. Lines are drawn to plot long words and short words
against the mean recall scores, which are repeated below.
Mean for short words at fast rate 7.5
Mean for short words at slow rate 3.75
Mean for long words at fast rate 3.75
Mean for long words at slow rate 6.25
The fact that the lines in Figure 18.1 cross over is an indication of the
interaction between the two IVs of short/long words and fast/slow
presentation rates. The variable of rate presentation affects the recall
scores on short and long words in opposite directions. More short words
are recalled with a fast rate and more long words at a slow rate.
Progress box eighteen

In two-way ANOVA (unrelated) there are two independent variables.
In two-way ANOVA (unrelated) the scores on four conditions for the
two IVs are produced by different participants.
Five sources of variance are tested by two-way ANOVA (unrelated):
between conditions on Variable A; between conditions on Variable B;
interaction between A and B (A B); error variance; and total
variance.
The appropriate parametric statistical test for analysing the ve
sources of variance is two-way ANOVA (unrelated).
Figure 18.1 Graph for 2 2 ANOVA (unrelated)
Two-way ANOVA (related) 19
19.1
In two-way ANOVA (related) the same participants do all the conditions.
Because each participant produces scores on all the conditions, the scores
for each participant are related. By calculating the differences between
participants related scores, it is possible to calculate individual variance
for participants.
In two-way ANOVA two independent variables (IVs) are tested. In two-
way ANOVA (related) the same participants do the two conditions on one
variable and also do the two conditions on the other variable.
An example would be an experiment in which the rst IV is a com-
parison between learning lists of long or short words. The second IV is a
comparison between fast and slow rates of presenting the lists of words,
as shown below.
Variable A (length of words)
Condition A
1
(short words)
Condition A
2
(long words)
Variable B (presentation rates)
Condition B
1
(fast rate)
Condition B
2
(slow rate)
In two-way ANOVA (related) the same participants will be doing the
two conditions on Variable A and also the two conditions on Variable B.
The same participants will be doing all four conditions listed in Table 19.1.
?
Suppose Variable A is a comparison between illustrated texts (A
1
) and Question 19.1
non-illustrated texts (A
2
). Variable B is a comparison between texts with a
title (B
1
) and texts with no title (B
2
).
Following the example in Table 19.1, ll in the four conditions on the
two variables, which are all done by the same participants.
Denitions of variances
19.2
Two-way ANOVA (related) indicates that there are two independent
variables.
For two-way ANOVA (related) there are several extra sources of
variance. The reason for this is that for a related design it is possible to
calculate individual variance between participants. This is based on the
differences between each participants related scores. Thus as many as
eight sources of variance can be calculated. These include:
Between conditions on Variable A (predicted)
Between conditions on Variable B (predicted)
Interaction between A and B (predicted)
Variance S (based on differences between participants related
scores)
Error A S (interaction between Variable A and S variance)
Error B S (interaction between Variable B and S variance)
Error A B S (three-way interaction between Variable A, Variable B
and S variance)
Total variance (including all the above sources of variance)
Table 19.1 Same participants doing all four conditions
Condition 1 Short words (A
1
) presented at a fast rate (B
1
)
Condition 2 Short words (A
1
2
)
Condition 3 Long words (A
2
) presented at a fast rate B
1
)
Condition 4 Long words (A
2
2
)
Condition 1
Condition 2
Condition 3
Condition 4
Note that it is universal in all statistical textbooks to use the term S (for
subjects) and not P for participants.
19.3
The answer for two-way ANOVA is two variables.
all four conditionis on two variables.
You will nd that if you answered the above three questions as suggested,
you will arrive at two-way ANOVA (related). This is the appropriate
parametric test for a two-way related design testing two independent
variables.
Using two-way ANOVA (related)
19.4
When to use
Two-way ANOVA (related) should be used for a related design when the
same participants are doing all conditions on two independent variables
and the data are interval.
Research hypothesis
It was predicted that there would be signicant effects of Variable A
between long and short words and Variable B between fast and slow rates
of presenting the word lists. No interaction was predicted between the two
variables.
Sample data
The same four participants are allocated to the four combinations of
Variable A conditions and Variable B conditions. This means that each
participant will have four scores in the four conditions, amounting to
16 scores in all. These scores and means are shown in Table 19.2.
Because there are so many sources of variances and interactions for
ANOVA (related) the data in Table 19.2 have to be shown in three separate
tables of cells. These tables represent summaries of the same data in
different combinations: A, B and A B interaction (Table 19.3); A, S
and A S interaction (Table 19.4); and B, S and B S interaction
(Table 19.5).
Notes on Table 19.3
1 This table has the same structure of Variable A and Variable B cells as
Table 18.3 in Chapter 18.
2 The totals are added for the scores in the four cells and the totals of A
and B, all adding up to the grand total of 77.
Table 19.2 Recall scores for two variables (related)
A
1
(short words) A
2
(long words)
Participants B
1
(fast rate) B
2
(slow rate) B
1
(fast rate) B
2
(slow rate)
1 7 7 3 5
2 5 6 1 3
3 6 8 2 5
4 4 9 2 4
Means 5.5 7.5 2 4.25
Table 19.3 AB summary table
A
1
A
2
Totals B (T
b
)
B
1
7 5 3 1
6 4 2 2
22 8 30
B
2
7 6 5 3
8 9 5 4
30 17 47
Totals A (T
a
) 52 25 77 Grand total
Notes on Table 19.4
1 This table shows the related scores for each participant (S) on the two
Variable A conditions.
2 The related scores for each participant are added as Totals S on the
right-hand side of the table.
3 The totals for A are the same as in Table 19.3.
4 All the scores in each summary table add up to the grand total of 77.
Table 19.4 AS summary table
A
1
A
2
Totals S (T
s
)
S
1
7 7 3 5
14 8 22
S
2
5 6 1 3
11 4 15
S
3
6 8 2 5
14 7 21
S
4
4 9 2 4
13 6 19
Totals A (T
a
Notes on Table 19.5
1 This table shows the related scores for each participant (S) are on the
two Variable B conditions.
2 The related scores for each participant for Totals S are the same as in
Table 19.4.
3 The totals for B are the same as in Table 19.3.
4 All the scores in each summary table add up to the grand total of 77.
Rationale
The aim of ANOVA is to compare ratios of variance. The between-
conditions variance for Variable A, Variable B and the interaction A B
represent predicted variances between conditions. Differences in scores
due to irrelevant variables represent error variance. The prediction is that
there will be relatively large differences between conditions compared with
error variance. If there are random differences between conditions, as
stated by the null hypothesis, the predicted differences between conditions
will be small in relation to error variance. In this case, the null hypothesis
cannot be rejected.
Table 19.5 BS summary table
B
1
B
2
Totals S (T
s
)
S
1
7 3 7 5
10 12 22
S
2
5 1 6 3
6 9 15
S
3
6 2 8 5
8 13 21
S
4
4 2 9 4
6 13 19
Totals B (T
b
Step-by-step instructions for calculating two-way ANOVA
(related). These are given in Box K.
Box K
Step-by-step instructions for calculating 2 2 F ratios (related)
Note the following symbols 1
(see Tables 19.319.5):
T
2
a
= sum of A squared
totals Table 19.3
T
2
a
= 52
2
+ 25
2
T
2
b
= sum of B squared
totals Table 19.3
T
2
b
= 30
2
+ 47
2
T
2
ab
= sum of AB squared
totals for individual
cells in Table 19.3
T
2
ab
= 22
2
+ 30
2
+ 8
2
+ 17
2
T
2
s
= sum of squared totals
for each participant
Table 19.4
T
2
s
= 22
2
+ 15
2
+ 21
2
+ 19
2
T
2
as
= sum of AS squared
totals for individual
cells in Table 19.4
T
2
as
= 14
2
+ 11
2
+ 14
2
+ 13
2
+ 8
2
+ 4
2
+ 7
2
+ 6
2
T
2
bs
= sum of BS squared
totals for the
individual cells in
Table 19.5
T
2
bs
= 10
2
+ 6
2
+ 8
2
+ 6
2
+ 12
2
+ 9
2
+ 13
2
+ 13
2
n = numbers of participants n = 4
a = number of
conditions for
variable A
a = 2
b = number of
conditions for
variable B
b = 2
(
x)
2
x)
2
= 77
2
(
x)
2
N
= constant to be
subtracted from all SS
=
77
2
16
= 370.5625
x
2
= sum of each
individual score squared
x
2
= 449
Calculate SS
A
2
=
52
2
+ 25
2
4 2
370.5625
T
2
a
nb

(
x)
2
N

= 416.125 370.5625
= 45.5625
Calculate SS
B
3
T
2
b
na

(
x)
2
N
=
30
2
+ 47
2
4 2
370.5625
= 388.625 370.5625
= 18.0625
Calculate SS
S
4
T
2
s
ab

(
x)
2
N
=
22
2
+ 15
2
+ 21
2
+ 19
2
4
370.5625
= 377.75 370.5625
= 7.1875
Calculate SS
AB
5
T
2
ab
n

(
x)
2
N
SS
A
SS
B
=
22
2
+ 30
2
+ 8
2
+ 17
2
4
370.5625 45.5625 18.0625
= 434.25 370.5625 45.5625 18.0625
= 0.0625
Calculate SS
AS
6
T
2
as
b

(
x)
2
N
SS
A
SS
S
=
14
2
+ 11
2
+ 14
2
+ 13
2
+ 8
2
+ 4
2
+ 7
2
+ 6
2
2
370.5625 45.5625 7.1875
= 423.5 370.5625 45.5625 7.1875
= 0.1875
Calculate SS
BS
7
T
2
bs
a

(
x)
2
N
SS
B
SS
S
=
10
2
+ 6
2
+ 8
2
+ 6
2
+ 12
2
+ 9
2
+ 13
2
+ 13
2
2
370.5625 18.0625 7.1875
= 399.5 370.5625 18.0625 7.1875
= 3.6875
Calculate SS
tot
8
x
2

(
x)
2
N

= 449 370.5625
= 78.4375
Calculate SS
ABS
9
SS
tot
SS
A
SS
B
SS
S
SS
AB
SS
AS
SS
BS
= 78.4375 45.5625 18.0625 7.1875
0.0625 0.1875 3.6875
= 3.6875
Notes on calculations in Box K
1 Note the denitions and numerical examples of the symbols and the
calculation of the constant, which will be subtracted from each SS.
2 The formulae for SS
A
, SS
B
and SS
AB
are familiar from Box J in Chapter
18. SS
A
is based on the two squared totals for A in Table 19.3. SS
B
is
based on the two squared totals for B in Table 19.3. SS
AB
is based on the
totals of the squared scores in the cells in Table 19.3.
3 SS
S
is a new source of variance based on the squared totals for T
s
in
Table 19.4.
Calculate the degrees of freedom 10 df
A
= 2 1 = 1
df
B
= 2 1 = 1
df
S
= 4 1 = 3
df
AB
= 1 1 = 1
df
AS
= 1 3 = 3
df
BS
= 1 3 = 3
df
ABS
= 1 1 3 = 3
df
tot
= 16 = 15
Divide each SS by df to 11 MS
A
= 45.5625
obtain MS MS
B
= 18.0625
MS
S
= 2.3958
MS
AB
= 0.0625
MS
AS
= 0.0625
MS
BS
= 1.229
MS
ABS
= 1.229
Calculate F ratios allocating 12
correct df to each F ratio
F ratio for MS
A
=
MS
A
MS
AS

F
1,3
= 729
F ratio for MS
B
=
MS
B
MS
BS

F
1,3
= 14.7
F ratio for MS
AB
=
MS
AB
MS
ABS

F
1,3
= 0.051
F ratio for MS
S
=
MS
S
MS
ABS

F
3,3
= 1.95
The completed ANOVA table 13
is shown in Table 19.6
4 SS
AS
is a new equation for error based on the squared totals of the cells
in Table 19.4.
5 SS
BS
is a new equation for error based on the squared totals of the cells
in Table 19.5.
6 SS
tot
is based on the squares of all the individual scores in Tale 19.3
minus the constant
7 SS
ABS
is error calculated by subtracting all the other calculated sums of
squares from SS
tot
.
8 Follow the instructions for calculating the rather complicated df for
each source of variance.
9 The F ratios are calculated by using the calculated error which is
appropriate for each source of variance. MS
A
is compared with AS
error, MS
B
with BS error, and MS
AB
with ABS error. MS
S
is also com-
pared with ABS error.
Tables G(1) and G(2) are the appropriate tables for all ANOVA tests. Table
G (1) gives the critical vales for p < 0.05. Table G (2) gives the critical
values for p < 0.01. The calculated F has to be equal to or larger than the
critical values. You need to use the degrees of freedom for all three
between-conditions variances for A, B, AB, and the error variances. You
can use the df from step 10 in Box K (in our example df
A
= 1 and df
AS
= 3;
df
B
= 1 and df
BS
= 3; df
AB
= 1 and df
ABS
= 3. What makes this easier is that
all the variances have the same df 1,3. So we only have to look up the
critical values for df = 1,3.
Table 19.6 Two-way ANOVA table (related)
Sources of variance
Sums of
squares
Degrees of
freedom
Mean
squares F ratios
Variable A (word length) 45.5625 1 45.5625 F
1,3
= 729
Variable B (presentation rate) 18.0625 1 18.0625 F
1,3
= 14.7
S (subjects) 7.1875 3 2.3958 F
3,3
= 1.95
A B (interaction) 0.0625 1 0.0625 F
1,3
= 0.051
Error AS 0.1875 3 0.0625
Error BS 3.6875 3 1.229
Error ABS 3.6875 3 1.229
Total 78.4375 15
In Table G(1) df for A, B, AB, is shown along the
1
row and df error is
shown down the
2
column. Locate 1 on
1
along the top line and follow
the
2
column down until you nd 3. Where they intersect is the critical
value of 10.13. The calculated F for A of 729 in Box K is much larger than
this critical value and it is also larger than the critical value in Table G (2)
for df = 1,3 of 34.1. So the null hypothesis for Variable A can be rejected
(p < 0.01). The calculated F for B of 14.7 is larger than the critical value of
10.13 in Table G (1) but smaller than the critical value of 34.1 in Table G
(2). So the null hypothesis for Variable B can be rejected (p < 0.05).
The calculated F for A B of 0.051 is obviously too small so the null
hypothesis for the A B interaction cannot be rejected. It is common
practice to test the signicance of the F for S variance, hoping that indi-
vidual differences between participants will not be a signicant source
of variance. In our example, the calculated F of 1.95 (df = 3,3) is smaller
than the critical value of 9.28 in Table G (1). So the differences between
participants (S) are not signicant.
Conclusions
If you look at the means in Table 19.2, you will nd that, as predicted,
signicantly more short words were recalled than long words and
more words recalled at a slow rate than a fast rate. These are main
effects of the two variables separately (see discussion of main effects in
Chapter 17, Section 17.3). There was no signicant interaction between
the two variables and the differences between participants were not
signicant.
Using the SPSS computer program
19.5
The calculation of so many different sources of error variance is much
more complicated for ANOVA (related). It is likely that that you may
have access to a computer program like SPSS to help you with the actual
calculations.
Inputs to SPSS computer statistical packages for ANOVA
(related)
The scores for the four conditions will have to be input to SPSS, indicating
that the scores come from the same participants. This will help you to
understand which scores from Tables 19.3, 19.4 and 19.5 you should be
inputting to the computer.
Outputs of SPSS computer packages for ANOVA (related)
The main output will be in the form of an ANOVA (related) table as in
Table 19.6.
The other output will be a graph showing the relationships between the
variables, as in Figure 19.1.
In Figure 19.1 fast rate and slow rate are printed at the bottom of the
graph and the short words and long words are plotted against the mean
recall scores (see the means in Table 19.2).
The two parallel lines in Figure 19.1 show that short words had bigger
mean recall scores than long words. However, there is no interaction
between the two variables of short and long words and fast and slow
presentation rates because neither variable has any effect on the other
variable.
Note about mixed designs
19.6
It is possible to use a mixed ANOVA design. What is meant by a mixed
ANOVA design is that it is a mixture between same participants and
Figure 19.1 Graph for 2 2 ANOVA (related)
There are still two variables in the experiment, Variable A and Variable
B. The difference is that the same participants do both conditions on
Variable A, while different participants do each condition on Variable B.
An example would be that the same four participants learn both
short words and long words (Variable A). This would control for possible
individual differences between participants learning ability.
Different groups of participants would be presented with the lists of
words at a slow rate or a fast rate (Variable B). Different participants
might be selected because participants might be affected by being exposed
to two different rates of presentation.
A mixed design is shown in Table 19.7.
Notes on Table 19.7
1 Six participants took part in this mixed design.
2 All of the six participants did both the A
1
and the A
2
conditions. They
produced a pair of related scores for short words and for long words.
3 Two different groups of three participants did the B
1
and B
2
conditions.
The scores for fast and slow rates were unrelated.
We have given you only a very brief introduction to mixed ANOVA with
a mixture of same participants and different participants on the two
independent variables. If you want to select a mixed ANOVA you will have
to ask advice. Alternatively, you can use the appropriate SPSS computer
program.
Table 19.7 Example of a mixed design
B
1
A
1
(short words)
A
2
(long words)
Totals
S (T
s
)
Totals
A(T
a
)
B
1
(fast rate) P
1
8 4 12
P
2
7 4 11
P
3
7 5 12
22 13 35
B
2
(slow rate) P
4
4 4 8
P
5
6 7 13
P
6
7 5 12
17 16 33
Totals B (T
b
) 39 29 68
Grand total
Progress box nineteen

In two-way ANOVA (related) there are two independent variables.
The scores in four conditions on the two variables are produced by
the same participants.
Eight sources of variance are tested by two-way ANOVA (related).
The ve main sources include:
Between conditions on Variable A
Between conditions on Variable B
Interaction A B (AB)
Individual variance (S)
Total variance
The other three variances are accounted for by interactions between
these variances in interaction with S variance, some of which consti-
tute different types of error.
Interaction A S (error AS)
Interaction B S (error BS)
Interaction A B S (error ABS)
The appropriate parametric statistical test for analysing these eight
sources of variance is two-way ANOVA (related).
It is possible to use a mixed design in which one variable is related
and the other variable is unrelated.
Relationships between variables Part IV
Chi-square 20
Comparison with experiments
20.1
In Chapter 1 we introduced the idea that all psychological research is
concerned with testing relationships between variables. In Chapter 2
experiments were introduced as a special case of relationships between
variables. Experimenters manipulate independent variables and predict
the outcome in terms of effects on measured dependent variables.
This makes it possible to predict differences between experimental
conditions.
All the statistical tests that were introduced in Parts II and III were
designed to test data from experiments. The non-parametric tests and the
parametric tests, including ANOVA, are all designed to test differences
between conditions.
In psychological research a researcher tests a research hypothesis about
relationships between variables. The only issue is how much control the
researcher has over the variables in the research design.
Measuring categories
20.2
The essential characteristic of chi-square is that it does not deal at all with
scores. Both variables are represented as categories. This means that for
chi-square the only measure of participants behaviour on both variables is
what categories participants are allocated to.
An example might be to observe whether participants are more likely to
be able to solve a problem if they are given a hint. The measure is how
many people can solve the problem with and without a hint.
For the chi-square test it is only appropriate to count the number of
participants who solve the problem and the number of participants
who do not solve the problem. It is absolutely crucial to understand that
Relationships between variables 144
the times taken to solve the problem are not relevant. The only relevant
measure is whether participants fall into the category of solving the prob-
lem or into the category of not solving the problem.
Nominal data
20.3
Because the number of participants in each category is the relevant
measure, this kind of measure is known as nominal data. The use of the
term nominal emphasizes that that this type of measurement can only
give a label, or name, to the categories. This is completely distinct from the
scores used in other kinds of research.
In our experiment the names (labels) given to categories are as follows:
Variable A Giving hints
Category 1 Giving a hint
Category 2 No hint
Variable B Solving problems
Category 1 Solving a problem
Category 2 Not solving a problem.
Another term that is sometimes used for nominal data is categorical
data. This term is used because data are measured as the number of
participants in each category.
?
A researcher gave one group of participants instructions about the Question 20.1
advantages of helping strangers and another group of participants
warnings about the possible dangers of helping strangers (Variable A). The
measure was counting the numbers of participants who were observed
helping a stranger or not helping a stranger (Variable B).
(a) What were the two categories for Variable A?
(b) What were the two categories for Variable B?
Predicting categories
20.4
The same participants are tested on both variables. The reason for this is
that the research hypothesis predicts that participants who are given a hint
will be able to solve a problem. In order to test this hypothesis it is essential
that the same participants are given a hint and then given the problem to
Chi-square 145
solve. This makes it possible for the researcher to make predictions about
the relationships between the two variables.
On the other hand, the pairs of conditions for each variable are dened
independently. It is impossible for the same participants to be given a
hint as well as not being given a hint. Equally, no one participant can
simultaneously be in the category of being able to solve the problem and
also in the category of not being able to solve the problem.
An example might be that a researcher predicts that participants who
are in the category of being given favourable instructions about helping
strangers are more likely to be in the category of helping an old lady
across a road. Participants who are in the category of being warned about
helping strangers are more likely to end up in the category of not helping a
stranger.
There is one other special feature that arises from predicting the
number of participants in different categories. In a typical experiment the
experimenter selects the experimental conditions and decides how many
participants will be used for each of the conditions.
The crucial distinction is that in a chi-square design the prediction
is formulated as the number of participants who will turn out to be in
each category. Rather than being a prior decision of the researcher, the
predicted numbers of participants in each category cannot be known until
the research is carried out.
In order to ensure that there will be a reasonable number of partici-
pants in each category, it is usual to recommend that quite a lot of
participants are used for chi-square. Otherwise it might turn out that,
in spite of the instructions, very few participants are prepared to help
other people.
20.5
Relationships between variables?
The answer to this question is Yes.
Are participants allocated to categories?
The answer is Yes.
If you look up relationships between variables and nominal categories
in the Decision Chart you will arrive at the chi-square test. This is the
appropriate statistical test for testing categorical data.
Using the chi-square test
20.6
When to use
The chi-square test should be used when participants are allocated to
categories and the data are nominal. Remember that the test only deals
with categories, not scores.
Research hypothesis
The researcher predicted that a higher proportion of technology
students would use regular study patterns, as compared with social science
students.
Sample data
A hundred participants are selected, one group of 50 social science
students and another group of 50 technology students. A questionnaire is
sent to all the students asking them to report whether their study patterns
fall into one of three categories:
1 Regular day-to-day study
2 Irregular bursts of study
3 Mixture of both kinds of study
There were 44 replies from social science and 42 replies from technology
students. The replies about study patterns were allocated to the three
categories. The results are shown in Table 20.1.
Chi-square 147
Notes on Table 20.1
1 The table shows how many students are allocated to the three categories
of study patterns.
2 The three study patterns are shown on the top line.
3 The two groups of social science and technology students are shown on
the left-hand side of the table.
4 There are six cells in the table for two groups of students and three study
patterns (2 3). These six cells are labelled cell 1, cell 2, etc.
5 In each cell the number of students in each category is indicated. In
cell 1, 6 social science students reported regular study patterns; in cell 2,
15 social science students reported irregular study patterns; in cell 4,
10 technology students reported regular study patterns; in cell 6, 24
technology students reported mixed study patterns.
6 These numbers are known as observed frequencies because the number
in each cell represents the frequency with which the students fall into the
three study pattern categories.
7 The totals for the three study patterns are shown at the bottom of the
table and the totals for the two groups of students on the right hand
side.
8 The grand total for all participants is 86, representing the 44 social
science students plus the 42 technology students.
Table 20.1 Reported study patterns
Study patterns Totals of
students
Regular Irregular Mixed
Group 1
(social
science
students)
Cell
1
6
Cell
2
15
Cell
3
23 44
Group 2
(technology
students)
Cell
4
10
Cell
5
8
Cell
6
24 42
Totals for
study
patterns
16 23 47 86 Total
number of
partici-
pants (N)
?
(a) How many social science students reported mixed study patterns in Question 20.2
cell 3?
(b) How many technology students reported irregular study patterns in
cell 5?
(c) What was the total for regular study patterns?
(d) What was the total for mixed study patterns?
(e) What was the total of technology students?
Rationale
Table 20.1 gives the observed frequencies of the two categories of students
who reported three categories of study patterns. These observed frequen-
cies have to be compared with expected frequencies. Expected frequencies
are calculated as if there were no bias towards a relationship between
the two sets of categories. The expected frequencies represent the null
hypothesis.
The expected frequencies have to be calculated taking into account the
relevant totals. These totals reect the proportions of the 44 social science
and 42 technology students who would be expected to be in the three study
pattern categories if it were equally probable that all students would be
found in all categories.
The observed frequencies in Table 20.1 are compared with the null
hypothesis of expected frequencies. If the observed frequencies are signi-
cantly different from the expected frequencies, the null hypothesis can be
rejected.
Step-by-step instructions for calculating chi-square
These are given in Box L.
Chi-square 149
Box L
Step-by-step instructions for calculating the value of
2
Note the observed 1
frequencies (O) in cells 16
in Table 9.1.
Cell 1: O = 6
Cell 2: O = 15
Cell 3: O = 23
Cell 4: O = 10
Cell 5: O = 8
Cell 6: O = 24
Calculate the expected 2
frequencies (E) for cells
Cell 1: E =
16 44
86
= 8.19
16 in Table 9.1 by
multiplying the two
Cell 2: E =
23 44
86
= 11.77
relevant marginal totals
for each cell and dividing
Cell 3: E =
47 44
86
= 24.05
by the total number of
participants (N = 86).
Cell 4: E =
16 42
86
= 7.81
Cell 5: E =
23 42
86
= 11.23
Cell 6: E =
47 42
86
= 22.95
Calculate the value of
2
3
using the formula
2
=
(O E)
2
E

Remember that squares of
minus differences result in
positive numbers.
=
(6 8.19)
2
8.19
+
(15 11.77)
2
11.77
+
(23 24.05)
2
24.05
+
(10 7.81)
2
7.81
+
(8 11.23)
2
11.23
+
(24 22.95)
2
22.95

= 0.58 (cell 1) + 0.89 (cell 2) + 0.05 (cell 3)
+ 0.61 (cell 4) + 0.93 (cell 5) + 0.05 (cell 6)
= 3.11
Calculate the degrees of 4
freedom:
c = number of columns c = 3
r = number of rows r = 2
df = (c 1) (r 1) df = (3 1) (2 1) = 2
Notes on calculations in Box L
1 The observed frequencies in each cell are taken from the cells in Table
20.1.
2 In order to calculate expected frequencies relevant totals have to be
identied in the table. The relevant totals are as follows: Cell 1 shows the
6 social science students who have regular study patterns. The relevant
totals for cell 1 are the 44 social science students and the 16 regular
study patterns reported by both types of students. For cell 5 the relevant
totals are the total of 42 technology students and the total of 23
students who reported irregular study patterns. Once the relevant totals
have been identied for each cell, the expected frequencies for each
cell are calculated by multiplying these totals and dividing them by the
grand total of 86 participants, as shown in Step 2.
3 The formula in step 3 calculates the sum of the squared differences
between observed and expected frequencies.
4 The df are calculated as the three cells for study patterns minus 1 (3
1 = 2) and the two cells for types of students minus 1 (2 1 = 1). These
are multiplied to produce df = 2.
?
What are the relevant totals for the following cells in Table 20.1? Question 20.3
(a) Cell 2
(b) Cell 3
(c) Cell 4
(d) Cell 6
Looking up the signicance of chi-square in Table E
You have already used Table E for the Friedman test (related) in Chapter
11 and for the KruskalWallis test (unrelated) in Chapter 12. In both these
cases, chi-square in Table E was used when there were large numbers of
participants.
The calculated value of chi-square has to be larger than the critical
values in Table E. To look up chi-square look along the row for df = 2. The
critical value for p < 0.05 is 5.99.
In Box L the calculated chi-square is 3.11. As this is smaller than the
critical value of 5.99 the null hypothesis cannot be rejected. Another way
of expressing this is that the observed frequencies were not signicantly
different from the expected frequencies.
Chi-square 151
Conclusions
The research hypothesis that more technology students would use regular
study patterns as compared with social science students was not supported.
The failure to nd signicant results can probably be explained by the
large number of students who reported mixed study patterns, as shown in
Table 20.1.
Chi-square can only test overall relationships between variables, which
is equivalent to a two-tailed hypothesis. The reason for this is that it is
possible to interpret relationships between variables in more than way. For
example, it is just as likely that students who have regular study patterns
might decide to study technology, rather than that studying technology
results in students learning good study habits.
Progress box twenty

Chi-square is based on predicted relationships between two variables
represented as categories.
Categories are identied for each variable.
The measured data are nominal, which involves counting the number
of participants in each of the categories.
To calculate chi-square observed frequencies are compared with
expected frequencies.
Pearson product moment correlation 21
Comparison with experiments
21.1
In all psychological research a researcher is testing a research hypothesis
about relationships between variables. The only issue is how much control
researchers have over the variables in the research design. Experimenters
select independent variables and measure the dependent variable of par-
ticipants scores. The research hypothesis predicts differences between
scores.
The Pearson product moment correlation is a test of correlation. Predic-
tions about correlations are quite different. In correlation the research
hypothesis predicts that two variables will be associated with each other.
Associations between variables are the basis for all correlations.
Measuring variables in correlations
21.2
The essential characteristic of correlational designs is that both variables
are represented as continuous ranges of scores. The measures of par-
ticipants performance on both variables produce a range of scores.
An example would be if children were given a spelling test and a reading
test. The measure of spelling would be a range of spelling scores produced
by the children and the measure of reading scores would be a range of
reading scores produced by the children.
The spelling scores would represent one variable and the reading scores
a second variable. It is important to note that in correlational designs the
same participants produce scores on both variables. This makes it possible
for the researcher to make predictions about relationships between scores
on both variables.
The fact that both variables are continuous has advantages. In experi-
ments it is necessary to manipulate categories of participants, groups of
good and less good spellers and groups of good and less good readers. The
advantage of a correlational design is that the full range of spelling scores
can be compared with the full range of reading scores.
Because the two variables in a correlation design are represented by
continuous ranges of scores, this counts as interval data. Numerical
calculations of sums of squares are required for the parametric Pearson
correlation test.
Predicting positive correlations
21.3
The researcher predicts a relationship between the two variables, in our
example childrens scores on spelling and reading. The prediction might be
that over the whole range of spelling ability and reading ability children
who are good spellers will also tend to score highly on a reading test, while
children at the bottom end of the spelling scale will not do so well at
reading.
In this case it is predicted that the scores on the two variables will be
positively related. This means that children who have high spelling scores
will also have high reading scores. Likewise, low spelling scores are pre-
dicted to be associated with low reading scores.
Because the variables move in the same direction, this is known as a
positive correlation. High scores on spelling go together with high scores
on reading, medium scores go with medium scores and low scores go with
low scores.
?
Which of the following are most likely to result in a high positive Question 21.1
correlation? Which are not likely to be correlated at all?
(a) Height and shoe size.
(b) Number of cinema tickets sold and the number of customers in the
audience.
(c) Amount of spinach eaten and wins on football pools.
It is obvious from the examples in Question 21.1 that in some cases
positive correlations between pairs of scores are likely to be high. Relation-
ships between other pairs of scores are likely to produce a much lower
amount of correlation, or no correlation at all. Another way of putting
this is that some correlations are very high, while others are low or
non-existent.
Predicting negative correlations
21.4
So far we have been talking about the amount of positive correlation
between pairs of scores, like spelling and reading scores. But what about
cases when high scores on one variable might be expected to be associated
with low scores on another variable? For instance, children who achieve
excellent scores on a quiz about sports may score low on writing essays,
and vice versa.
It is important to realize that, if high scores on one variable are associ-
ated with low scores on another variable, this is still a correlation, but it is
in the opposite direction to a positive correlation. This is known as a
negative correlation because the scores on one variable are moving in the
opposite direction to scores on the other variable.
In our example it is predicted that high scores on the sport quiz will be
associated with low scores on essay writing. This is a negative correlation
between high scores on one variable and low scores on the other variable.
?
Which of the following are likely to be positive correlations or negative Question 21.2
correlations?
(a) Temperatures in winter and size of electricity bills.
(b) Amount of rain and sale of umbrellas.
Correlation coefcients
21.5
The next question is how to measure precisely the amount of correlation
between two variables.
Correlations are measured in terms of correlation coefcients,
which indicate the size of a correlation between two variables. Correlation
coefcients run from zero, i.e. no correlation, to +1 for a perfect positive
correlation. A perfect negative correlation is represented by 1.
In psychological research it is unlikely that any scores would be so
perfectly correlated. Because of the innite variability of human beings,
there are sure to be some good spellers who are bad at reading. For this
reason, most correlations between psychological variables are not perfectly
correlated. Most correlations fall somewhere between the extremes of
+1 (positive), 1 (negative) and zero (none).
?
(a) Which of the three correlation coefcients listed below expresses the Question 21.3
lowest and the highest correlations? List them in order from lowest to
highest.
+0.5 0 0.9
(b) Which of these correlation coefcients is most likely to express the
relationship between number of miles travelled and the cost of a single
second class ticket?
(c) Which of these correlation coefcients is most likely to express the
relationship between the number of pedestrian crossings in a town
and average earnings?
(d) Which of these correlation coefcients is most likely to express the
relationship between the amount of time spent practicing the piano
and the number of football games played?
To sum up, a correlation coefcient of zero means that there is no
relation at all between the variables. The higher the correlation coefcient,
whether positive or negative, the higher the correlation. Smaller corre-
lation coefcients (such as +0.2 and 0.2) indicate lower correlations
between variables.
21.6
Is it predicted that two variables are correlated?
The answer is Yes.
If you look up relationships between variables and correlations in the
Decision Chart you will arrive at the Pearson product moment correlation
test. This is the appropriate statistical test for testing correlations between
two variables.
Using Pearson correlation product moment correlation
21.7
When to use
The Pearson product moment correlation should be used when correlated
scores are measured and the data are interval.
Research hypothesis
The prediction is that high arithmetic scores will be positively correlated
with high reading scores. This is a one-tailed hypothesis predicting a posi-
tive correlation between the reading and arithmetic varibles.
?
Suppose a researcher predicted that there would either be a positive or a Question 21.4
negative correlation between two sets of measured scores. Would this be
a one-tailed or a two-tailed hypothesis?
Sample data
An arithmetic test and a reading test were administered to 12 children. The
arithmetic scores (Variable A) and the reading scores (Variable B) are
shown in Table 21.1.
Notes on Table 21.1
1 The arithmetic and reading scores are shown for each of the 12
participants.
2 The sums are shown for the scores on both variables.
Table 21.1 Scores on arithmetic test and reading test
Parti-
cipants
Variable A
(arithmetic
score)
Variable B
(reading
score) a b a
2
b
2
1 32 17 544 1 024 289
2 54 13 702 2 916 169
3 68 14 952 4 624 196
4 93 10 930 8 649 100
5 87 16 1392 7 569 256
6 24 7 168 576 49
7 49 6 294 2 401 36
8 35 18 630 1 225 324
9 97 19 1843 9 409 361
10 62 13 806 3 844 169
11 44 9 396 1 936 81
12 73 12 876 5 329 144
Total

a= 718

b= 154
(a b) = 9533
a
2
= 49 502

b
2
= 2174
3 In the third column each of the a scores is multiplied by each of the b
scores (a b) and the sum of these is calculated. This represents the
predicted variance between the A and B variables.
4 In the if column each a score is squared and summed and in the sixth
column each b score is squared and summed. These summed squares
represent the total variance.
Rationale
The aim of the Pearson correlation is to test whether scores on one
variable are related to scores on the other variable. This is measured by
whether a correlation coefcient is high enough to be signicant. It is
important to note that a correlation has to be reasonably high to be con-
sidered as being possibly signicant. If a researcher is investigating a very
large number of subjects, a quite low correlation might turn out to be
signicant. However, unless there is a reasonably high correlation in the
rst place, the correlation may be signicant but actually quite low. It is
usually accepted that there should be a reasonably high correlation, say
0.5 or better, before carrying out a test to assess its signicance.
The correlation is a ratio between the predicted variance in and the total
scores as a ratio between variance in scores. If the ratio is low, the null
hypothesis cannot be rejected. The ratio is denoted by r.
Step-by step instructions for calculating r
There are given in Box M.
Box M
Step-by-Step Instructions for Calculating the Value of r
For each participant multiply the A 1
variable score by the B variable score
See a b column in Table 21.1.
Square individual scores for Variable 2
A and for Variable B
See a
2
and b
2
columns in Table 21.1
Add the totals few all columns: 3
N = number of participants N = 12
a b = total of a b column

a b = 9533
a and
b = totals for each variable

a = 718
b = 154
Notes on the calculations in Box M
1 Steps 1, 2 and 3 refer back to the calculations carried out in Table 21.1.
2 The formula in step 4 gives the calculation of the predicted variance on
the top line of the formula.
3 The sums of squares on the bottom line of the formula give the variance
for the A variable multiplied by the variance for the B variable. These
two variances represent total variance. Note that the square root has to
be taken of the total variance.
4 The total formula in Step 4 represents the ratio between the predicted
variance and total variance.
5 In step 5 the df are dened as number of participants minus 1 for
Variable A and number of participants minus 1 for Variable B.
This amounts to total number of participants minus 2 (in our example
12 2 =10 df).
(
a)
2
and (
b)
2
= totals for each
variable squared
(
a)
2
= 718
2
(
b)
2
= 154
2
a
2
and

b
2
= sums of squared
individual scores
a
2
= 49 502
b
2
= 2174
Find the value of r from the formula 4
r =
N
a b

a

b
[N
a
2
(
a)
2
][N
b
2
(
b)
2
]
=
(12 9533) (718 154)
(12 49 502 718
2
)(12 2174 154
2
)
=
114 396 110 572
(594 024 515 524)(26 088 23 716)
=
3824
78 500 2372
=
3824
186 202 000
=
3824
13 645.5
= 0.28
Calculate df, i.e. number of 5
participants minus 2
df = N 2 = 12 2 = 10
Looking up the signicance of r in Table H
As was pointed out in the rationale, a correlation measured by r has to be
at least as high as 0.5 to be considered to be high enough to look up its
signicance. The calculated r in Box M of 0.28 is a pretty low correlation.
Normally, it would be assumed that the correlation is not signicantly
different from a zero correlation and so the null hypothesis cannot be
rejected.
However, in order to demonstrate how to look up r in Table H we
will tell you how to do this. Table H gives the critical values for different
levels of signicance levels for one-tailed and two-tailed hypotheses. The
calculated r has to be equal to or larger than the critical values. You
rst need to identify the degrees of freedom (in our example df = 10).
If you look along the row for 10 df for a one-tailed signicance level
(p < 0.05) the critical value is 0.4973. The calculated r of 0.28 is lower
than this value and so, as we expected, the null hypothesis cannot be
rejected.
Conclusions
The r of 0.28 was too low to be signicant. The research hypothesis that
there would be a positive correlation between arithmetic and reading
scores was not supported.
Note that, even if the value of r had been signicant, there are usually
several possible explanations for a relationship between variables. It is not
always possible to identify the precise reason for the relationship.
Possible explanations for a relationship between two variables of
reading scores and spelling scores are shown below:
1 Spelling ability may be responsible for improving reading.
2 Being a good reader improves spelling.
3 Both good spelling and good reading are the result of some other vari-
able altogether, such as a positive attitude by parents towards school
lessons.
?
(a) What would be three possible explanations if a signicant positive Question 21.5
correlation was found between arithmetic and reading? Hint: Look at
the example of three explanations for spelling and reading above.
(b) What would be a likely explanation for a correlation between a cock
crowing and an alarm clock going off ?
In some cases it may be likely that a correlation represents a causal
explanation. The strong and consistent correlation between smoking and
lung cancer has led to acceptance of a causal connection. But, in principle,
it is possible that there is some other common factor affecting the use of
cigarettes and a predisposition to lung cancer. That is why medical
research has been undertaken to discover physiological mechanisms in
order to indicate a direct causal link.
Progress box twenty one

Correlation represents a predicted relationship between two
variables.
Correlations are measured on a scale running from +1 (perfect
positive correlation) to 0 (no correlation) to 1 (perfect negative
correlation).
A reasonably high correlation (0.5) is required before testing the
signicance of a correlation.
Introduction to simple linear
22
regression
Comparisons with correlation
22.1
The type of linear regression discussed in this chapter is called simple
regression. In simple linear regression only two variables are tested, as in
correlation. This is to distinguish it from multiple regression, which will be
described in Chapter 23.
There are quite a few similarities between correlation and simple linear
regression. The main point that they have in common is that they are both
concerned with relationships between two variables.
Another similarity is that the scores on both variables represent con-
tinuous scores. The scores for both variables are interval data.
We will show that both correlation and simple linear regression can
be demonstrated using graphs that are called scatterplots. We will also see
that error is measured in a very similar way for both types of analysis.
The big difference is the aim of the researcher. In Chapter 21 we set out
to show how two sets of scores are correlated with each other. For
example, Both spelling scores and reading scores are measured and the
correlation between then is calculated. The researcher is simply interested
in whether a correlation is found between the two variables. In linear
regression, by contrast, the scores on the two variables are not treated as
correlated. Instead the scores on one variable are used to predict scores on
the other variable.
An example would be if childrens reading scores are used to predict
their arithmetic scores. The assumption would be that, if the researcher
knows a childs reading score it would be possible to predict their
arithmetic score.
Predictor variable and criterion variable
22.2
The variable that is selected to predict scores on the other variable is
usually called a predictor variable. The other variable, on which scores are
to be predicted, is often called the criterion variable.
It is usual to denote the scores on the predictor variable by the symbol X
and the scores on the predicted criterion variable by the symbol Y. In our
example the predictor variable X represents the reading scores, which are
used to predict the variable Y arithmetic scores.
?
(a) Is Variable A (arithmetic scores) or Variable B (reading scores) the Question 22.1
predictor variable?
(b) Which of these two variables is the criterion variable?
(c) Which of the scores on the two variables is represented by the symbol
X and which by the symbol Y?
We will take as an example the scores for two variables given in
Table 22.1. The aim of the researcher is to predict childrens arithmetic
scores on the basis of their reading scores.
Table 22.1 Arithmetic scores (Variable A)
and reading scores (Variable B)
Child Arithmetic score Reading score
1 9 2
2 10 3
3 12 1
4 6 1
5 11 4
6 9 1
7 12 5
8 16 8
9 13 5
10 10 3
11 13 6
12 14 7
Introduction to simple linear regression 163
Scatterplots
22.3
Scatterplots can be drawn to show relationships between two variables.
They are called scatterplots because dots representing scores on variables
are plotted in graphs, enabling a scatter to be seen.
Scatterplots can be used to represent a correlation between scores. As
we shall see, they can also be used to show the relationship between the
predictor variable and the predicted criterion variable.
The scatterplot for the arithmetic and reading scores in Table 22.1 is
shown in Figure 22.1.
A scatterplot is a graph constructed by placing the criterion variable
(arithmetic scores) up the vertical axis of the graph and predictor variable
(reading scores) along the horizontal axis at the bottom of the graph.
Figure 22.1 Scatterplot arithmetic and reading scores from Table 22.1
Scatterplot of arithmetic and reading scores from Table 22.1.
Each child is allocated a dot in the scatterplot, which represents that
childs scores for arithmetic and reading. For example, child 1 in Table 22.1
has an arithmetic score of 9 and a reading score of 2. These scores are
plotted up the vertical axis as a score of 9 and along the bottom axis as a
score of 2.
In order to locate this dot on the scatterplot look along the bottom line
for a reading score of 2 and from the vertical axis for an arithmetic score
of 9. Where these two scores meet is the dot for child 1. You will notice
that there is only one dot for a reading score of 2 and that dot has an
arithmetic score of 9. So this dot represents the reading and arithmetic
scores for child 1.
?
Which child numbers are represented by the rst three dots at the left-hand Question 22.2
side of the scatterplot? Hint: Look at the scores in Table 22.1.
The points plotted in the scatterplot in Figure 22.1 seem to show that
there is a positive linear relationship between arithmetic scores and
reading scores. If you look at the reading scores along the bottom line,
high reading scores tend to predict high arithmetic scores and lower
reading tend to predict lower arithmetic scores.
Regression equation for calculating a regression line
22.4
As we said earlier, in simple linear regression one of the two variables is a
predictor variable, which is used to predict scores on the criterion variable.
The predictor variable X (reading scores) is used to predict variable Y
(arithmetic scores). The predicted Y scores are denoted by Y.
The aim is to construct a line on the scatter plot which links up the
predictor variable scores (X) and the predicted Y scores (Y). This is called
a regression line, and shows how the Y scores are predicted by the X
scores.
The regression equation denes the regression line. The regression
equation is given by.
Y = A + BX.
A is the point at which Y is calculated when the X variable is set to zero. It
is called the intercept because it is where the regression line cuts the Y line
in the absence of the X predictor variable.
The equation Y = A + BX is a mathematical expression for a straight
line. This line is called the regression line because it represents the effect of
the predictor X scores on the predicted Y scores.
The slope of the regression line represents the degree to which the
scores on Y are predictable by the scores on the predictor variable X.
This make it possible to read off the predicted scores for Y for each score
of X.
The reason for dening A is that it provides a baseline when the pre-
dictor X scores are zero. This is a technical concept. There is often no
psychological meaning attached to considering what a childs arithmetic
score would be if they had a nil reading score.
An example to indicate what regression line connecting the dots in the
scatterplot in Figure 22.1 could look like is shown in Figure 22.2.
Figure 22.2
Residual error
22.5
The straight regression line in Figure 22.2 represents a linear relationship
between the dots in the scatterplot representing predictor X reading scores
and predicted Y arithmetic scores.
The regression line can be thought of as an ideal condition in which all
the predicted Y scores should be the same as the actual arithmetic scores.
The dots in the scatterplot represent the actual reading and arithmetic
scores.
It is important to note that not all the dots representing scores are
connected by the regression line. This is due to individual differences
between participants. It would not be expected that all the children would
have exactly the same range of arithmetic and reading scores.
For example, in Table 22.1 child 3 obtained a score of only 1 for reading
but a score of 12 for arithmetic. This is the dot on the left-hand side of the
scatterplot representing a score of 1 on reading and a score of 12 on
arithmetic.
This particular example goes against the prediction that arithmetic
scores can be predicted on the basis of reading scores. Looking at the low
reading score for child 3, a low arithmetic score would have been predicted
rather than a high score of 12.
The fact that the regression line does not connect all the dots means that
there is error in the dots representing childrens scores. Error is measured
as the differences between the actual scores and the scores predicted by the
regression line.
These differences can be thought of as deviations in scores from the
predicted scores. These deviations are called residual error. All error is
residual because error is what is left over after predicted variance has been
accounted for.
Deviations from the regression line can be calculated as differences
between predicted Y scores and actual observed Y scores. The deviations
between predicted and actual Y scores are measured as Y Y. These
deviations represent residual error.
It is interesting to note that some of the dots are quite near the regres-
sion line. The aim is to solve the linear regression equation so as to result in
the smallest deviations between predicted Y scores and observed Y scores.
Rationale for using ANOVA as a test of linear regression
22.6
It may sometimes be useful to be able to read off Y scores on the basis of X
scores in order to predict new cases of arithmetic scores. However,
researchers are often more interested in whether regression lines represent
a good prediction of the predicted Y scores in relation to observed Y
scores.
The deviations between predicted Y scores and actual Y scores are
measured by the sums of deviations squared. This is how sums of squares
are calculated in ANOVA as deviations between the scores and the
means.
So it should not come as too much of a surprise that ANOVA is the
appropriate test for the signicance of the predictions made on the basis
of predictor variables.
ANOVA is used to test different sources of variance. In linear regression
there are two sources of variance, which make up total variance.
1 Predicted variance in Y scores accounted for by the predicted Y
scores on the basis of the predictor variable X (predicted linear
regression).
2 Residual error as a result of unpredicted differences in Y scores not
explained by the predictor variable X (residual error variance).
Sums of squares are calculated for each source of variance. In the usual
way these sums of squares are divided by the df to produce mean squares.
The predicted mean square divided by the error mean square results in the
F ratio.
If the ANOVA is not signicant this implies that there is too much
residual error reecting deviations from the regression line. If all the
dots are scattered randomly, as stated by the null hypothesis, it will be
impossible to calculate a best-t regression line to minimize deviations. In
this case the null hypothesis cannot be rejected.
ANOVA table to test linear regression
22.7
An ANOVA table for simple linear regression based on the data in Table
22.1 is shown in Table 22.2.
Table 22.2 ANOVA table for linear regression
Source df Sum of squares Mean squares F ratio Signicance
Regression 1 57.49 57.49 27.69 (df 1,10) p < 0.0005
Residual 10 20.76 2.08
Total 11 78.25
Notes on Table 22.2
1 The sources of variance are regression (predicted) and residual (error).
2 There were 12 participants so the total df is 12 1 = 11. The df for
regression is df = 1 because there is only one predictor variable. The df
for error are the remaining df 11 1 = 10.
3 The residual sum of squares is calculated by subtracting the regression
sum of squares from total sum of squares.
4 A large amount of variance is accounted for by the predicted regression
with a relatively small error.
5 This results in a highly signicant F ratio.
Conclusions
The signicance of ANOVA for linear regression demonstrates that
reading scores are a good predictor of arithmetic scores.
It is important to note that this prediction is based on a relationship
between variables. For example, it might be possible that the linear relation-
ship between variables works in the opposite direction. It might turn
out that it is easier to predict childrens reading scores (Y variable) from
knowing their arithmetic scores (predictor X variable).
Progress box twenty two

In linear regression the scores on a predictor variable (X) are used to
predict scores on a criterion variable (Y).
Regression lines are calculated to minimize the deviations between
the predicted Y scores and the observed Y scores (residual error).
Linear regression is tested by ANOVA with regression as the pre-
dicted variance and deviations as residual error.
Multiple regression 23
Comparisons with simple regression
23.1
In Section 21.5 we described correlation coefcients as a measure of the
correlational association between two variables. As explained in Section
22.1, the difference between correlation and regression is that in regression
one of the variables is selected as a predictor variable, which predicts
scores on the other variable. The association between the predictor
variable and the criterion variable is measured as a single correlation
coefcient.
Multiple regression introduces several predictor variables. For instance,
there could be three predictor variables all predicting scores on the pre-
dicted criterion variable. Because the same participants produce scores on
all four variables, it is possible to calculate correlation coefcients between
all the variables.
As in Chapter 22, in our discussion of multiple regression we will be
using the terminology of predictor variables and criterion variable. As
with simple linear regression, the predictor variables are denoted X, in
this case three predictor variables X
1
, X
2
and X
3
. The criterion variable is
denoted Y.
The aim of multiple regression analysis is to investigate the extent to
which the scores on the predictor X variables succeed in predicting scores
on the Y variable.
Objectives of multiple regression
23.2
This section will give a brief introduction to some of the main features of
multiple regression in order to prepare you for later discussions.
Objective 1
To assess the extent to which scores on predictor X variables are associated
with Y scores.
One of the main methods to test this is to carry out correlation
analysis between all the variables to assess the signicance of correlations
between all pairs of variables, each predictor with the predicted criterion
variable and also between pairs of predicators. Multiple correlations
are usually shown in the form of a multiple correlation matrix, as in
Table 23.1.
Notes on Table 23.1
1 Each of the gures in the table represents a correlation between each
of the three predictors (X
1
, X
2
, X
3
) and the predicted criterion variable
(Y). The matrix also shows the correlation coefcients between the
predictors themselves.
2 You should notice that the correlations between variables are given
twice. For instance, the correlation between Y and X
1
of 0.67 is the same
as the correlation between X
1
and Y of 0.67.
3 The correlation of a variable with itself, e.g. Y and Y, is always 1.
?
Question 23.1
(a) What is the correlation between Y and X
2
?
(b) What is the correlation between X
2
and X
2
?
It is important to realize that the most desirable outcome is to nd
signicant correlations between the scores on the predictor X variables and
the scores on the criterion Y variable.
In Table 23.1 the correlations between the three X variables and the Y
variable are all high (0.67, 0.52, 0.74). These correlations between the
Table 23.1 Sample multiple correlation matrix
Y X
1
X
2
X
3
Y 1 0.67 0.52 0.74
X
1
0.67 1 0.42 0.21
X
2
0.52 0.42 1 0.52
X
3
0.74 0.21 0.52 1
predictor X variables and the Y criterion variable would be good evidence
that the scores on the predictor X variables have a signicant effect on the
scores on the Y variable.
The correlations in the matrix can be thought of as individual corre-
lations, each of which could be tested by the Pearson correlation test of r.
To express the effect of all the correlations together the term multiple R
is used.
Objective 2
To assess the signicance of the variance on the predicted criterion
variable contributed by all the predictor variables simultaneously.
It is possible to calculate the square of multiple R, called R
2
or R-square.
R
2
is very important because it is a measure of the predicted multiple
regression that can be explained by the predictor variables. This is
expressed as the amount of total variance in Y scores accounted for by the
predictor X variables.
The next step is to assess the signicance of the predicted variance. The
way in which the regression variance is tested is to compare it with error
variance. This results in a ratio between predicted regression variance and
error variance.
The idea of a ratio between two sources of variance, consisting of
predicted variance and error variance, should remind you of Chapters 14
and 15, which described ANOVA in just these terms. It is not surprising,
then, that ANOVA is used to test R
2
as a measure of predicted regression
variance. An outline ANOVA table showing the sources of variance is
shown in Table 23.2
Notes on Table 23.2
1 Under the rst column, sources of variance, the line for regression
variance represents the proportion of predicted variance (measured by
R
2
) contributed by the predictor X variables.
Table 23.2 Outline ANOVA table
Sources of variance Sum of Squares df Mean Square F Sig.
Regression
Residual
Total
2 Residual represents the proportion of error variance in Y scores that is
not accounted for by the predictor variables. This is called residual error
because, as with all error, it represents what is left over after all the
predicted variance has been accounted for.
3 Residual error can be dened formally as deviations between the pre-
dicted Y scores and the actual observed Y scores (Y Y), as discussed
in Section 22.5. In practice, residual error is calculated by subtracting
predicted regression variance from total variance, as is usual in ANOVA
tables.
Objective 3
To assess the predictive effects of individual X predictor variables
So far we have considered the effects of the variance in Y scores
accounted for by all X predictor variables.
It is also important to estimate the contribution of each X predictor
variable in predicting the Y scores. In order to compare the effects of
individual predictors we need to standardize each predictor variable so
that they can be compared on an identical scale. These standardized
versions of the B coefcients for each predictor variable are known as beta
values. Betas will be described in a later section.
As you will remember from Chapters 9 and 10, t tests are designed to
test comparisons between pairs of conditions. This is why t tests are par-
ticularly suitable for making individual comparisons between coefcients
for each predictor variable and the error associated with each one.
Types of multiple regression
23.3
There are three main types of multiple regression analysis, namely,
standard multiple regression, sequential multiple regression and statistical
multiple regression. The last two types were evolved to meet specic
problems related to the order in which each predictor variable enters the
multiple regression analysis.
We need to assess which type of multiple regression should be used
after inspection of the data. It is usual to start with standard multiple
regression.
Standard multiple regression
In standard multiple regression all predictors are entered simultaneously
into the multiple regression. This allows the effect of each predictor vari-
able to be compared with the effects of all the other predictor variables at
the same time.
Technically, so that each predictor variable can be tested against all the
other predictor variables, each predictor variable is treated as if it has
been entered last into the analysis. The point of this as if condition is
that each variable will be analysed in relation to all the variables that
have already been entered. This meets the requirement that each predictor
variable is compared with all the previously entered variables.
What standard multiple regression analysis does is to determine the
effect of all predictor variables in contributing to total variance, i.e.
the proportion of variance that can be attributed to all the predictor
variables.
As its name implies, standard multiple regression should always be used
as a norm.
Sequential multiple regression
In standard multiple regression the predictor variables are entered so as
to ensure that all variables are analysed against all the other variables. But
sometimes the researcher may want to decide on a particular order of the
predictor variables being entered into the multiple regression.
This can be achieved by using sequential multiple regression. Note that
sequential regression is sometimes called hierarchical regression because it
determines the hierarchical order in which variables are entered into the
multiple regression.
In this type of multiple regression the researcher determines the order in
which each predictor variable enters the analysis. The effect of each pre-
dictor variable in contributing to predicted variance is assessed separately
at its point of entry. This contrasts with standard multiple regression,
which analyses the combined effect of all the predictor variables.
The researcher determines the order of entry of each predictor variable
on the grounds of theoretical evidence about the relative weight of each
single variable. The researcher gives priority to variables that have been
shown to have a signicant effect on predicted performance.
One problem is that researchers might favour certain variables as
opposed to others variables that could have an equal effect. In order to
control for this possibility of bias, researchers are recommended to select
an order of entry in which less favoured predictor variables are entered
into the analysis before the theoretically based predictor variables. If the
theoretical predictors are still found to have an effect, this is good evidence
in support of these selected variables.
Statistical multiple regression
There are various sub-types of statistical multiple regression. The order of
entry of the predictors is based purely on statistical grounds, leading often
to grave problems of interpretation of results. All types of statistical
regression should only be applied to very large samples when a researcher
wants to nd a subset of predictors that explain a large proportion of
predicted variance,
Statistical multiple regression is a controversial, and somewhat
advanced, technique that should be applied with great care in special
circumstances.
23.4
Are Y scores predicted by X scores on predictor variables?
The answer is Yes.
If you look up relationships between variables and predictor variables
you will arrive at multiple regression. This consists of procedures for
analysing regression for a set of multiple predictors.
Using multiple regression
23.5
When to use
Multiple regression is used when we want to predict effects of several
predictor variables on a criterion variable. Predictor variables may be
measured on a continuous scale or be categorical. An example of a
categorical variable would be a predictor variable that noted whether
children had attended a nursery school or not attended a nursery school.
This is called a categorical variable because there are only two outcomes,
attendance or not attendance.
Note that the criterion variable must always be a continuous variable.
Research hypothesis
It is predicted that three predictor variables, peer interaction, child
teacher interaction and supervision of daily activities, will affect scores on
the criterion variable of group involvement.
Sample data
In the example, group involvement (GI) is the predicted criterion variable
Y and peer interaction (PI), childteacher interaction (CTI) and super-
vision of daily activities (SDA) are the three predictor variables. All
three predictor variables when tested independently by other researchers
were found to be signicantly correlated with group involvement. Thirty
children attending a crche and their teachers were the participants in this
study.
Researchers carrying out the study used observational techniques to col-
lect data on the four variables. You will notice from the scores in Table 23.3
Table 23.3 Table of scores
Children GI PI CTI SDA
1 81.11 3 2 2
2 94.00 4 4 4
3 85.20 5 7 5
4 95.57 6 7 7
5 81.40 3 3 2
6 87.50 4 2 3
7 76.61 6 4 2
8 82.60 2 2 1
9 83.13 3 4 4
10 90.21 4 4 4
11 85.00 4 2 3
12 93.10 4 4 2
13 81.25 5 2 2
14 83.55 4 2 2
15 88.50 3 2 1
16 84.37 3 1 2
17 91.12 6 6 4
18 84.60 4 4 4
19 85.72 4 4 4
20 91.62 7 4 6
21 74.93 3 3 2
22 85.52 4 4 2
23 88.53 5 7 2
24 92.19 6 6 7
25 82.38 6 6 4
26 89.00 3 3 2
27 88.66 4 1 4
28 87.84 3 2 2
29 88.59 4 6 1
30 89.61 4 4 4
that the scores on the variables were measured on different rating scales.
For example GI scores are expressed with decimal points and all appear to
be very high in comparison with the scores obtained for the other three
variables. This demonstrates researchers often use different rating scales
for measuring variables. All observational sessions were video-taped so
that independent judges could validate the scores originally allocated by
the researchers.
Rationale
The rationale for multiple regression is twofold. In the rst place it is
necessary to determine whether the regression variance, i.e. the variance
accounted for by the effect of the predictor variables on the predicted
criterion variable, is signicantly higher than the variance due to the effect
of residual error. As explained in Section 23.2 (Objective 2), regression
variance is tested by ANOVA.
Secondly, it is necessary to determine the relative weight of each of
the individual predictors on the predicted criterion variable scores. As
explained in Section 23.2 (Objective 3), regression analysis performs a
series of t tests between each predictor and its own error to test the signi-
cance of individual predictor variables.
If the residual variance is high, as predicted by the null hypothesis, the
null hypothesis cannot be rejected.
Step-by-step instructions for interpreting multiple regression
There are no calculations in this section. The aim is to understand
computer printouts of multiple regression data.
The equation for calculating regression lines in multiple regression is
Y = A + B
1
X
1
+ B
2
X
2
+ B
3
X
3
Here Y represents the predicted scores on the criterion variable of
group involvement (GI); A is the intercept, representing the score on Y
when the three X variables are zero; B
1
X
1
is the regression coefcient
for the slope of the regression line for the predictor variable PI multi-
plied by values of PI; B
2
X
2
is the regression coefcient for the slope of
the regression line for the predictor variable CTI multiplied by
values of CTI; and B
3
X
3
is the regression coefcient for the slope of
the regression line of the predictor variable SDA multiplied by values
of SDA.
As in Chapter 22, Section 22.4, the multiple regression equation is a
mathematical expression for a straight line representing the slope of the
regression line.
As in the simple equation, Y equals predicted Y scores on the criterion
variable GI.
In multiple regression A represents a common origin point for all three
predictor variables. For this reason, A is dened as the intercept for all
predictor variables. A is also known as the constant, because it is the same
for all three predictor variables.
In the multiple regression equation there are several BX coefcients. The
regression equation for multiple regression predicts Y scores by calculating
regression lines for each of the three predictor variables.
The next steps interpret the outputs of a simplied SPSS computer
analysis of multiple regression. Table 23.4 shows the output of the SPSS
program in the form of descriptive statistics. It gives the means, the
standard deviations and the number of participants for the predicted
criterion variable (GI) and the three predictor variables (PI, CTI and
SDA).
Table 23.5 shows the correlations between variables in the form
of a multiple correlation matrix. It also gives the signicance of each
correlation.
Table 23.4 Descriptive statistics
Mean Std. Deviation N
GI 86.4468 4.88375 30
PI 4.20 1.215 30
CTI 3.73 1.799 30
SDA 3.13 1.634 30
Table 23.5 Multiple correlation matrix
GI PI CTI SDA
Correlation GI 1.000 0.299 0.349 0.490
PI 0.299 1.000 0.625 0.646
CTI 0.349 0.625 1.000 0.505
SDA 0.490 0.646 0.505 1.000
GI
Sig. (1-tailed) 0.054 0.029 0.003
PI 0.054 0.000 0.000
CTI 0.029 0.000 0.002
SDA 0.003 0.000 0.002
Notes on Table 23.5
1 The multiple correlation matrix is in the same form as that shown in
Table 23.1.
2 Correlation coefcients are shown between the variables including the
predicted criterion variable (GI) and the three predictor variables (PI,
CTI and SDA).
3 Using the Pearson product moment correlation (see Chapter 21), sig-
nicance can be calculated for each correlation.
Looking at the correlations in Table 23.5, there are signicant corre-
lations between the predictor variables CTI and SDA with the criterion
variable (GI). But the correlation between PI and GI is just above the
conventional p < 0.05 and so is not signicant.
There are also some signicant correlations between predictor variables,
between SDA and PI and SDA and CTI. Signicant correlations between
predictor variables can indicate that scores on one predictor variable are
highly associated with scores on another predictor variable. For example,
family income and choice of private schools might be so highly correlated
that they do not add anything extra to the analysis. It can be argued that
there is no need to investigate income if this also determines parents
choice of schools.
In Table 23.5 there may seem to be quite high signicant correlations
between predictor variables. It is important to note that, if there were very
high correlations between a pair of predictor variables, the computer pro-
gram itself would automatically throw out one of the paired variables. As
none of the correlations between predictor variables have been identied
as being too high, all the predictor variables are retained in the multiple
regression for further analysis.
The data in Table 23.6 give the values of several important measures
required for multiple regression.
Notes on Table 23.6
1 Multiple R represents a summary of all the correlations in the multiple
correlation matrix.
2 R
2
represents the proportion of variance contributed by the predictor
variables.
Table 23.6 Calculation of
multiple R and R
2
Multiple R 0.511
R
2
0.261
Adjusted R
2
0.176
Std. error 4.43409
3 Adjusted R
2
is calculated taking into count the number of predictors. It
should be used when many predictor variables are included in the
regression. It is a corrective measure because, when many predictors are
added to the regression, R
2
can become articially high just by chance.
The next step in the multiple regression analysis involves nding out
whether the variance accounted for by the three predictors on the pre-
dicted criterion variable is signicant when compared to the effect of
residual error.
The value of the constant A has to be tested as one of the predictors that
may affect Y scores.
In Section 23.2 (Objective 2), the use of ANOVA was introduced as a
test of the proportion of variance accounted for by the regression co-
efcients calculated for each predictor variable, including the A constant.
The ANOVA table shown in Table 23.7 is based on the data in Table
23.3.
Notes on Table 23.7
1 Regression variance is based on the proportion of variance in Y scores
accounted for by all the predictor variables, including A, as measured by
R
2
.
2 Residual variance represents the error in Y scores which is not
accounted for by predictor variables.
3 Total variance represents all the variance in Y scores.
4 Residual error is calculated by subtracting predicted regression variance
from total variance.
5 The total df are calculated as 30 1 = 29. 3 df are allocated to the three
predictor variables. The df for residual error are the remaining df = 26.
6 The F ratio is calculated in the usual way.
7 Looking up Table G(1) for ANOVA, the critical value of F for df 3, 26 is
2.98. The calculated F of 3.06 is larger than the critical value so the null
hypothesis can be rejected (p < 0.05). In Table 23.7 there is an equivalent
signicance level of 0.046.
8 The conclusion is that the predicted regression variance accounted for
by the predictor variables is supported by the data.
Table 23.7 ANOVA test of predictor variables
Sum of Squares df Mean Square F Sig.
Regression 180.489 3 60.163 3.060 0.046
Residual 511.190 26 19.661
Total 691.679 29
The last step in the multiple regression analysis is to calculate the rela-
tive weights of each individual predictor variable in predicting scores on
the criterion Y variable. The terms relevant to this objective are shown in
Table 23.8.
Notes on Table 23.8
1 The rst column of B coefcients are the same B coefcients describing
the slope of each regression line, plus the constant shown in our regres-
sion equation above.
2 The standard error in the next column is a measure of the error for the B
coefcients for each predictor variable.
3 The nal column indicates the t values found for the constant and the
three predictor variables, i.e. the comparisons between the Bs and their
associated error terms.
4 The standardized beta coefcient for each predictor variable gives
the relative weight of each predictor variable on the unit of change in
the criterion Y variable.
Using the data in Table 23.8, the method for calculating the relative
weight of each predictor variable is to standardize the B coefcients so
that they can be compared on an identical scale. The standardized B
coefcients are called beta coefcients. Because betas are standardized,
beta coefcients allow us to test the relative contribution of each predictor
variable in terms of a unit change in the scores of the criterion variable.
Using these standardized beta coefcients, each predictor variable can be
compared in terms of how much it contributes to units of change in the
criterion variable compared with other variables.
The method for comparing individual predictor variables is to test the
contribution of each variable to the total variance in Y scores. You will
see in Table 23.8 that, for each of the B coefcients and the constant, a
standard error is given.
Table 23.8 Tests for relative weights of individual variables
Unstandardized
coefcients
Standardized
coefcients
B Std. Error Beta t Sig.
(Constant) 82.216 2.979 27.596 0.000
PI 0.506 0.996 0.126 0.508 0.616
CTI 0.507 0.595 0.187 0.852 0.402
SDA 1.424 0.670 0.477 2.126 0.043
a Dependent Variable: GI
As explained in Section 23.2 (Objective 3), t tests are used to test the
signicance of coefcients against their own error. In this case each B
coefcient is tested against the appropriate standard error. The ratio for
the t test is the B coefcient over its own standard error. Looking at the
signicance of the t tests the only predictor variable that was signicant
was SDA.
The other variable that was very signicant was the constant. You will
remember that the constant is A, which is the intercept for the regression
lines for all the predictor variables. The importance of the constant is that
it ensures that the same point of intersection holds for the population from
which the childrens scores drawn. In the t test the constant is compared
against the overall error of all predictor variables.
Conclusions
The multiple regression matrix in Table 23.5 shows the signicant correl-
ations between the four variables, GI (the predicted criterion variable) and
PI, CTI and SDA (the predictors). PI was not signicantly correlated with
the predicted criterion variable, making doubtful its contribution to scores
on GI. Correlations of the other two predictors with GI were signicant,
as were some of the correlations between predictors. None of the corre-
lations between predictor variables was identied as being too high. So
all the predictor variables were retained in the multiple regression analysis.
(It was earlier noted that when correlations between predictor variables
are too high the computer program automatically excludes one of the
variables.)
Multiple R was 0.511. R
2
was calculated as 0.261, representing the pre-
dicted proportion of variance, and adjusted R
2
was 0.176. The ANOVA to
assess the regression effects of all three predictor variables was signicant
(p < 0.05). The t tests to test the signicance of the B coefcients associ-
ated with each of the predictor variables conrmed that the only predictor
with a signicant contribution to the variance on the criterion variable
is SDA. This would indicate that the correlation coefcient for SDA
accounted for most of the differences in Y scores. This was borne out by
the high correlations between SDA and GI in the correlation matrix in
Table 23.5.
The interpretation of these results shows that supervision of daily
activities had a signicant effect on the selected criterion variable of group
involvement. Contrary to other ndings, group involvement and peer
interaction were not signicantly correlated. On the other hand,
childteacher interaction and supervision of daily activities, as predicted
by other studies, were signicantly correlated with group involvement. The
important point to make is that by means of regression we were able to
determine that, when these two predictors are compared simultaneously,
only one of them (SDA) appears to contribute signicantly to the variance
on the predicted criterion variable GI.
Why is peer interaction not correlated with group involvement? Group
involvement measures the extent to which children get involved in a task
set by the teacher and cooperate among themselves to get it done. The
amount of peer interaction among them may not inuence directly their
involvement in the task, as is borne out by the low correlation between PI
and GI. Other variables may be more responsible for their involvement,
namely CTI, which measures the type of interaction between teachers and
children, and SDA, which measures the quality of supervision that
teachers exert over daily activities, including tasks. In our example the
type of SDA had by far the strongest effect. In other words, children left
on their own with poor teacher supervision may not succeed in nishing
a task. When teachers possess good supervision qualities children are
more likely to get positively involved among themselves to achieve any
particular task.
Progress box twenty three

In multiple regression the scores on predictor variables (X
1
, X
2
, etc.)
are used to predict scores on a criterion variable (Y).
Regression lines are calculated by identifying the coefcients (B)
which dene the slopes for each predictor variable, and the point of
intersection common to all regression lines (A).
In multiple regression the proportion of variance contributed by all
predictor variables is compared with residual error using ANOVA.
The signicance of each predictor variable when compared to
its standard error is tested individually using t tests for each
comparison.
Standardized beta coefcients allow the relative weights of predictor
variables in predicting unit changes on the Y scores to be compared
on an identical scale.
General linear model 24
Introduction
24.1
The aim of this nal chapter is to introduce a general model that will bring
together the main types of statistical analysis. This overall model is called
the general linear model (GLM).
The term general simply means that the model applies equally to
regression analysis and to the analysis of variables in ANOVA.
Turning to the term Linear, the basic notion is that this refers to the
nature of the lines that are used in the statistical models we have con-
sidered so far. The assumption of linearity is that the lines are straight
lines. We will remind you that the regression equations described in
Chapters 22 and 23 were mathematical expressions that dene straight
lines. In multiple regression the regression coefcients dening the
regression slopes are based on regression lines, which are straight lines.
If you look back to Chapter 17, which introduces two-way ANOVA,
you will see that the relationships between variables are also based on the
slopes of straight lines. In the graph in Figure 17.1 the different slopes of
the straight lines for good readers and less good readers demonstrate that
there is an interaction between the two variables.
In Figure 17.2 the fact that the slopes of the straight lines are identical
indicates that there is no interactive relationship between the two variables.
In both multiple regression and ANOVA the assumption is that
relationships between variables are dened by the slopes of straight lines.
This is what is meant by a linear model.
The role of error in GLM
24.2
All statistics in psychology are based on a comparison between predicted
variance and error variance, expressed as a ratio between them. Error
represents the unpredicted variability in human behaviour, which has not
been predicted by the researcher.
The results of a research project are more likely to be signicant if
predicted variance is high and unpredicted error relatively low.
Because the GLM is a general model that accounts for different types of
statistical analysis, it always includes the predicted relationships between
variables and also an error term.
There are considerable similarities in the way that the error term in
GLM is calculated for multiple regression and ANOVA. In both cases, the
error term is calculated as error variance. Error variance is also called
residual error. This is because error is what is left over after the variance
due to predicted variables has been accounted for.
Application of the GLM to linear regression
24.3
The GLM has a lot in common with regression equations for multiple
regression lines given in Chapter 23. This accounts for the slopes of
straight lines representing scores predicted by selected predictor variables.
In multiple regression the variables are often continuous variables, like
a numerical rating scale. But it is possible to incorporate a categorical
variable, such as attending a nursery school.
Application of GLM to ANOVA
24.4
ANOVA is also based on relationships between variables expressed as
the slopes between straight lines. Ratios are calculated between predicted
variance and error variance.
In ANOVA the independent variables have to be categories. Examples
would be the two categories of being given a reading scheme or not being
given a reading scheme.
Another telling similarity between the two models is that ANOVA is
used as a test of the effects of variables in regression. It has been argued
that there is no essential difference between the predictor variables in
multiple regression and the independent variables in ANOVA. The degrees
of freedom depend in both cases on the number of predictor variables and
the number of independent variables.
Conclusions
24.5
We end this last chapter by identifying some lessons we hope that you will
have learned from this book. One lesson is that all types of research are
equally valuable.
General linear model 185
In t tests and ANOVA the research hypothesis predicts differences
between experimental conditions selected by the researcher. Selecting
experimental conditions to test a research hypothesis has the advantage that
the causal effects of several variables can be investigated simultaneously in
ANOVA.
In linear regression certain variables are selected as predictor variables.
It may be statistically possible that other variables could be regressed in
order to achieve the best-tting regression line. Exploring a research
hypothesis predicting relationships between variables has the advantage
that many interesting naturally occurring variables can be investigated
simultaneously in multiple regression.
This brings us full circle to the rst chapter in the book. Understanding
a research hypothesis allows the selection of statistical tests to fall into
place.
By now you will be able to nd your way easily around the Decision
Chart. The main distinction is whether a research hypothesis predicts
differences between experimental conditions or relationships between
variables.
There are a few other questions about related designs (same par-
ticipants) or unrelated designs (different participants); the number of
conditions and variables; and how data are measured.
This is all you need to know to select an appropriate statistical test.
Now that you have completed a variety of tests, we hope that statistics will
seem a far easier task than you may have thought when you rst started
out.
Answers to questions
Question 2.1
(a) The independent variable is learning lists of common and rare words.
(b) The dependent variable is recalling words.
(c) If there are higher recall scores in Experimental condition 1, this supports the
research hypothesis that it will be easier to recall lists of common words.
Question 4.1
The research hypothesis predicted that participants would have better recall scores
for simple texts. The higher mean score for ideas recalled in the simple text in
Condition 1 supports the research hypothesis that more ideas will be recalled from
simple texts.
Question 4.2
(a) In Table 4.1 three participants scored 5 in Condition 1 and one participant in
Condition 2.
(b) In Condition 1 the lowest score was 3 and the highest score was 10.
(c) In Condition 2 the lowest score was 1 and the highest score was 7.
Question 4.3
p < 0.001 represents the smallest probability of random scores because there is only
1 in a 1000 probability that that the scores are random.
Question 5.1
A rank of 5 has been allocated to the score of 7. The reason is that this is the next
highest score after 6, which has been allocated a rank of 4.
Question 6.1
(a) The difference for participant 3 is 2 and the difference for participant 4 is 3.
(b) The mean for Condition 1 is 24 4 = 6 and the mean for Condition 2 is 3.
Question 6.2
This is a one-tailed hypothesis because it makes a prediction in one direction that
students will recall more words when they are listening to music.
Question 6.3
(a) The mean scores are shown in Table A.1.
(b) The blank columns are completed in Table A.1. Because ()3 is the smaller
rank total, W = 3.
N = 8 1 (tie) = 7.
(c) In Table A, looking along the N = 7 row the calculated W = 3 is smaller than 4
but larger than 2. So the one-tailed hypothesis that texts with short sentences
will be easier to comprehend is supported (p < 0.05).
Table A.1 Wilcoxon test (Question 6.3)
Scores (1) Scores (2) d
Ranks
of d
Signed ranks
(plus)
Signed ranks
(minus)
7 8 1 1 1
10 6 +4 +4.5 +4.5
13 4 +9 +7 +7
8 4 +4 +4.5 +4.5
7 7 0 (tie)
8 10 2 2 2
6 3 +3 +3 +3
10 3 +7 +6 +6
Means 8.63 5.63 Ranks totals (+)25 ()3
Question 7.1
It is not possible to calculate differences between the two conditions because
groups of different participants are doing only one condition each.
Question 7.2
One group of different participants would see the text with a title and
another group of different participants would see the text without a title.
If the same participants saw both texts, they would nd it impossible to
forget the title in the no title condition.
Question 7.3
(a) Look at the overall ranks of scores in Table A.2. Check that the lowest rank of
1 is allocated to the lowest score of 1 and that the highest rank of 12 is
allocated to the highest score of 20.
(b) The rank totals are shown separately for each condition.
(c) The lower rank total for the left-hand condition indicates that lower
(i.e. quicker) times are taken to recognize words on the left-hand side of the
screen. This supports the research hypothesis.
Question 8.1
(a) 5 5 = 25
(b) 5 5 = 25
(c) 1 1 = 1
(d) 3 3 = 9
Table A.2 MannWhitney test (unrelated)
Left-hand condition Right-hand condition
Scores Ranks Scores Ranks
9 10 8 9
3 3 5 4.5
2 2 7 8
6 6.5 20 12
1 1 11 11
5 4.5 6 6.5
Rank totals T
1
= 27 T
2
= 51
Question 9.1
For Condition 1 the mean is 28 4 = 7, for Condition 2 it is 7.5.
Question 9.2
The hypothesis is one-tailed because it predicts in one direction that more words
will be recalled from a simple text.
Question 9.3
All the answers to (a)(e) are shown in Table A.3.
Question 10.1
The two scores of 5 do not represent tied scores because the two scores come from
different groups of participants.
Question 11.1
Note that for participant 5 the highest score of 7 is in Condition 2 so this is
assigned the highest rank of 3. For participant 6 the two scores of 3 are assigned
tied ranks of 1.5.
Table A.3 t test (related) (Question 9.3)
Participants Condition 1 Condition 2 d d
2
1 6 2 4 16
2 7 1 6 36
3 8 3 5 25
4 10 4 6 36
5 8 3 5 25
6 8 2 6 36
7 5 7 2 4
8 3 4 1 1
Means 6.88 3.25 d = 29 d
2
= 179
Conditions 1 Condition 2 Condition 3
5 3 1 7 3 5 2
6 3 1.5 3 1.5 6 3
Question 11.2
(a) For Condition 1 the mean is 8 4 = 2, for Condition 2 it is 5, and for
Condition 3 it is 5.5.
(b) These means indicate that fewer rare words were recalled in Condition 1 than
common or very common words in the other two conditions.
Question 11.3
(a) For Condition 1 the mean is 3.6, for Condition 2 it is 6.6, for Condition 3 it
is 7.
(b) The independent variable is the colour in which the prose passages were
printed. The dependent variable is the number of ideas recalled from each
passage.
(c) The Friedman test for related designs is appropriate because the same ve
subjects read all three prose passages.
(d) The horizontal ranks across the three conditions for each participant are
shown in Table A.4.
(e) The rank totals for each condition are shown in Table A.4.
Question 12.1
(a) The higher rank total for Condition 3 indicates that more very common words
were recalled than rare words in Condition 1.
(b) For Condition 1 the mean is 1.5, for Condition 2 it is 5, and for Condition 3 it
is 7.5. These means also indicate that more very common words (Condition 3)
were recalled than rare words (Condition 1).
Question 13.1
(a) The missing score is 16.
(b) The df are 4 1 = 3. This reects the fact that three out of the four scores in
Table 13.3 can vary. Given three scores and the total, the fourth score is pre-
dictable and so cannot vary.
Table A.4 Friedman test (Question 11.3)
Scores (1) Ranks Scores (2) Ranks Scores (3) Ranks
4 1 5 2 6 3
2 1 7 2.5 7 2.5
6 1.5 6 1.5 8 3
3 1 7 3 5 2
3 1 8 2 9 3
T
1
= 5.5 T
2
= 11 T
3
= 13.5
Question 13.2
(a) Only two participants had a score of 1. This is quite an extreme score in the
normal distribution.
(b) Five participants had a score of 6, which is quite near the centre of the
distribution of scores.
Question 15.1
The three presentation rates should be presented in different orders. The aim is to
counterbalance the order of conditions when all the conditions are presented
to the same participants.
Question 17.1
The totals of 9 and 16 represent the main effect of Variable B: good readers (2 +
7 = 9) and less good readers (8 + 8 = 16).
Question 18.1
Group 1 Illustrated texts (A
1
) with title (B
1
).
Group 2 Illustrated texts (A
1
) with no title (B
2
).
Group 3 Non-illustrated texts (A
2
) with title (B
1
).
Group 4 Non-illustrated texts (A
2
) with no title (B
2
).
Question 18.2
(a) A
1
/B
1
9867
A
1
/B
2
4335
A
2
/B
1
5334
A
2
/B
2
7567
(b) 30 + 15 + 15 + 25 = 85
45 + 40 = 85
Question 19.1
Condition 1 Illustrated texts (A
1
) with title (B
1
).
Condition 2 Illustrated texts (A
1
) with no title (B
2
).
Condition 3 Non-illustrated texts (A
2
) with title (B
1
).
Condition 4 Non-illustrated texts (A
2
) with no title (B
2
).
Question 20.1
(a) The two categories for Variable A were being given favourable instructions or
danger warnings about helping strangers.
(b) The two categories for Variable B were the number of participants who were
observed helping a stranger or not helping stranger.
Question 20.2
(a) In cell 3,23 social science students reported mixed study patterns.
(b) In cell 5,8 technology students reported irregular study patterns.
(c) The total for regular study patterns was 16.
(d) The total for mixed study patterns was 47.
(e) The total of technology students was 42.
Question 20.3
(a) The relevant totals for cell 2 are the total of 44 social science students and the
total of 23 students who reported irregular study patterns.
(b) The relevant totals for cell 3 are the total of 44 social science students and
the total of 47 students who reported mixed study patterns.
(c) The relevant totals for cell 4 are the total of 42 technology students and the
total of 16 students who reported regular study patterns.
(d) The relevant total for cell 6 are the total of 42 technology students and
the total of 47 students who reported mixed study patterns.
Question 21.1
(a) There is usually a high positive correlation between the height of a person
and the size of their shoes. This is not a perfect correlation because some tall
people have very small feet.
(b) There should be a perfect positive correlation between these variables because
the number of tickets sold should exactly correspond to the number of
customers (unless, of course, some people manage to slip into the cinema
without paying!).
(c) There is most unlikely to be any correlation between the two variables of
amount of spinach eaten and the amount of money won on the pools
unfortunately!
Question 21.2
(a) In winter, the lower the temperature, the higher electrical bills are likely to be.
This is a negative correlation because low temperatures go with high electricity
bills. The variables of temperature and the amounts charged for electricity are
moving in opposite directions.
(b) There is likely to be a positive correlation between the amount of rain falling
and the number of umbrella sales.
Question 21.3
(a) The correlations should be listed in the order. 0, 0.5, 0.9.
(i) 0 is the lowest correlation coefcient because it indicates that there is no
correlation at all.
(ii) +0.5 is next highest because +0.5 is exactly half-way between 0 and 1.
(iii) 0.9 is the highest correlation coefcient because it is only slightly less
than a perfect negative correlation of 1.
(b) There is a positive correlation between number of miles in a journey and the
cost of a railway ticket. But this will not be a perfect correlation because on
some lines tickets can cost more or less for different length journeys (+0.5)
journeys.
(c) There is probably a zero correlation between number of pedestrian crossings
and average earnings.
(d) There is likely to be a highly negative correlation between practising the piano
and time available for football (0.9).
Question 21.4
If a researcher predicted that there would be a correlation, but that the correlation
might turn out to be either positive or negative, this would be a two-tailed
hypothesis making predictions in two directions.
Question 21.5
(a) Three possible relationships between reading and arithmetic are shown in
Figure A.5.
(b) It is pretty obvious that there is no direct relationship between a cock crowing
and an alarm clock going off. They are explained by another variable (early
rising).
Question 22.1
(a) Reading scores (Variable B) is the predictor variable.
(b) Arithemetic scores (Variable A) is the predicted criterion variable.
(c) It is a convention to refer the predictor variable (Variable B) as X and the
predicted criterion variable (Variable A) as Y.
Question 22.2
The three dots on the left of the scatter plot represent children who all obtained
very low reading scores of 1 but rather different arithmetic scores. If you look at
Table 22.1 you can see that child 4 has a score of 1 for reading combined with a low
score of 6 for arithmetic. Child 3 is the odd one out. A low score of 1 for reading
would be expected to go with a low arithmetic score. But child 3 has one of the
highest arithmetic scores of 12.
Question 23.1
(a) The correlation between Y and X
2
equals 0.52.
(b) The correlation between X
2
and X
2
equals 1. This is because a variable always
has a correlation of 1 with itself in the multiple correlation matrix.
Figure A.1 Relationship between variables (Question 21.5)
Recommended reading
There are a few books which provide some useful insights into experi-
mental design and statistics.
One general book is Research Methods and Statistics in Psychology by
Hugh Coolican. This book gives an introduction to designing experiments
and other types of research. Ethics is taken into consideration and there
are useful hints about describing data and writing up research reports. The
statistics are quite simple, but there are other compensations. Another
book which covers much of the same ground is Designing and Reporting
Experiments by Peter Harris.
Siegels book, Nonparametric Statistics for the Behavioral Sciences, is a
classic book on non-parametric tests. In fact, it was Siegel who made these
tests so popular.
There are also more advanced statistic textbooks. These tend to assume
quite a lot of knowledge and tend to be couched in mathematical
language. They contain calculations for the Tukey test for looking at com-
parisons between ANOVA conditions which is described in Chapter 16.
Two recent books are Doing Quantitative Psychological Research by David
Clark-Carter and Data Analysis for Psychology by George Dunbar.
Finally, there are many manuals for computerized statistical packages.
One of the most commonly used programs is SPSS, which can be used on
either PC or Macintosh.
References
Brace, N., Kemp, R. and Snelgar, R. (2003) SPSS for Psychologists. Basingstoke:
Palgrave Macmillan.
Clark-Carter, D. (1997) Doing Quantitative Psychological Research: From Design to
Report. Hove: Psychology Press.
Coolican, H. (1999) Research Methods and Statistics in Psychology (3rd edition).
London: Hodder & Stoughton.
Dunbar, G. (1998) Data Analysis for Psychology. London: Arnold.
Harris, P. (2002) Designing and Reporting Experiments in Psychology (2nd edition).
Maidenhead: Open University Press.
Pallant, J. (2001) SPSS Survival Manual. Maidenhead: Open University Press.
Siegel, S. (1956) Nonparametric Statistics for the Behavioral Sciences. New York:
McGraw-Hill.
Statistical tables
Table A Critical values of W at various levels of probability (Wilcoxon)
The statistic W denotes the smaller sum of total signed ranks. For any N (number of participants minus
ties) the observed value of W is signicant at a given level of signicance if it is equal to or less than the
critical values shown in the table.
Level of signicance for one-tailed test Level of signicance for one-tailed test
0.05 0.025 0.01 0.005 0.05 0.025 0.01 0.005
Level of signicance for two-tailed test Level of signicance for two-tailed test
N 0.10 0.05 0.02 0.01 N 0.10 0.05 0.02 0.01
5 1 28 130 117 102 92
6 2 1 29 141 127 111 100
7 4 2 0 30 152 137 120 109
8 6 4 2 0 31 163 148 130 118
9 8 6 3 2 32 175 159 141 128
10 11 8 5 3 33 188 171 151 138
11 14 11 7 5 34 201 183 162 149
12 17 14 10 7 35 214 195 174 160
13 21 17 13 10 36 228 208 186 171
14 26 21 16 13 37 242 222 198 183
15 30 25 20 16 38 256 235 211 195
16 36 30 24 19 39 271 250 224 208
17 41 35 28 23 40 287 264 238 221
18 47 40 33 28 41 303 279 252 234
19 54 46 38 32 42 319 295 267 248
20 60 52 43 37 43 336 311 281 262
21 68 59 49 43 44 353 327 297 277
22 75 66 56 49 45 371 344 313 292
23 83 73 62 55 46 389 361 329 307
24 92 81 69 61 47 408 379 345 323
25 101 90 77 68 48 427 397 362 339
26 110 98 85 76 49 446 415 380 356
27 120 107 93 84 50 466 434 398 373
* Dashes in the body of the table indicate that no decision is possible at the stated level of signicance because the
numbers of participants are too small.
Table B Critical values of U at various levels of probability (MannWhitney)
For any n
1
and n
2
the observed value of U is signicant at a given level of signicance if it is equal to or
less than the critical values shown in Tables B(1) and B(2)
Table B(1) Critical values of U for a one-tailed test at 0.005; two-tailed test at 0.01*
n
1
n
2
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
1 0 0
2 0 0 0 1 1 1 1 2 2 2 3 3 3 4 4 4
3 0 0 1 2 2 3 3 4 5 5 6 7 7 8 9 9 10 11
4 0 1 2 3 4 5 6 7 8 9 10 11 12 14 15 16 17 18
5 0 1 2 4 5 6 8 9 11 12 13 15 16 18 19 20 22 23 25
6 0 2 3 5 7 8 10 12 14 16 17 19 21 23 25 26 28 30 32
7 0 2 4 6 8 11 13 15 17 19 21 24 26 28 30 33 35 37 39
8 1 3 5 8 10 13 15 18 20 23 26 28 31 33 36 39 41 44 47
9 1 3 6 9 12 15 18 21 24 27 30 33 36 39 42 45 48 51 54
10 1 4 7 11 14 17 20 24 27 31 34 37 41 44 48 51 55 58 62
11 1 5 8 12 16 19 23 27 31 34 38 42 46 50 54 57 61 65 69
12 2 5 9 13 17 21 26 30 34 38 42 47 51 55 60 64 68 72 77
13 2 6 10 15 19 24 28 33 37 42 47 51 56 61 65 70 75 80 84
14 2 7 11 16 21 26 31 36 41 46 51 56 61 66 71 77 82 87 92
15 3 7 12 18 23 28 33 39 44 50 55 61 66 72 77 83 88 94 100
16 3 8 14 19 25 30 36 42 48 54 60 65 71 77 83 89 95 101 107
17 3 9 15 20 26 33 39 45 51 57 64 70 77 83 89 96 102 109 115
18 4 9 16 22 28 35 41 48 55 61 68 75 82 88 95 102 109 116 123
19 0 4 10 17 23 30 37 44 51 58 65 72 80 87 94 101 109 116 123 130
20 0 4 11 18 25 32 39 47 54 62 69 77 84 92 100 107 115 123 130 138
Table B(2) Critical values of U for a one-tailed test at 0.01; two-tailed test at 0.02*.
n
1
n
2
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
1
2 0 0 0 0 0 0 1 1
3 0 0 1 1 1 2 2 2 3 3 4 4 4 5
4 0 1 1 2 3 3 4 5 5 6 7 7 8 9 9 10
5 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
6 1 2 3 4 6 7 8 9 11 12 13 15 16 18 19 20 22
7 0 1 3 4 6 7 9 11 12 14 16 17 19 21 23 24 26 28
8 0 2 4 6 7 9 11 13 15 17 20 22 24 26 28 30 32 34
9 1 3 5 7 9 11 14 16 18 21 23 26 28 31 33 36 38 40
10 1 3 6 8 11 13 16 19 22 24 27 30 33 36 38 41 44 47
11 1 4 7 9 12 15 18 22 25 28 31 34 37 41 44 47 50 53
12 2 5 8 11 14 17 21 24 28 31 35 38 42 46 49 53 56 60
13 0 2 5 9 12 16 20 23 27 31 35 39 43 47 51 55 59 63 67
14 0 2 6 10 13 17 22 26 30 34 38 43 47 51 56 60 65 69 73
15 0 3 7 11 15 19 24 28 33 37 42 47 51 56 61 66 70 75 80
16 0 3 7 12 16 21 26 31 36 41 46 51 56 61 66 71 76 82 87
17 0 4 8 13 18 23 28 33 38 44 49 55 60 66 71 77 82 88 93
18 0 4 9 14 19 24 30 36 41 47 53 59 65 70 76 82 88 94 100
19 1 4 9 15 20 26 32 38 44 50 56 63 69 75 82 88 94 101 107
20 1 5 10 16 22 28 34 40 47 53 60 67 73 80 87 93 100 107 114
Table C Critical values of t at various levels of probability (t test)
For any particular df the observed value of t is signicant at a given level of
signicance if it is equal to or larger than the critical values shown in the table.
Level of signicance for one-tailed test
0.10 0.05 0.025 0.01 0.005 0.0005
Level of signicance for two-tailed test
df 0.20 0.10 0.05 0.02 0.01 0.001
1 3.078 6.314 12.706 31.821 63.657 636.619
2 1.886 2.920 4.303 6.965 9.925 31.598
3 1.638 2.353 3.182 4.541 5.841 12.941
4 1.533 2.132 2.776 3.747 4.604 8.610
5 1.476 2.015 2.571 3.365 4.032 6.859
6 1.440 1.943 2.447 3.143 3.707 5.959
7 1.415 1.895 2.365 2.998 3.499 5.405
8 1.397 1.860 2.306 2.896 3.355 5.041
9 1.383 1.833 2.262 2.821 3.250 4.781
10 1.372 1.812 2.228 2.764 3.169 4.587
11 1.363 1.796 2.201 2.718 3.106 4.437
12 1.356 1.782 2.179 2.681 3.055 4.318
13 1.350 1.771 2.160 2.650 3.012 4.221
14 1.345 1.761 2.145 2.624 2.977 4.140
15 1.341 1.753 2.131 2.602 2.947 4.073
16 1.337 1.746 2.120 2.583 2.921 4.015
17 1.333 1.740 2.110 2.567 2.898 3.965
18 1.330 1.734 2.101 2.552 2.878 3.922
19 1.328 1.729 2.093 2.539 2.861 3.883
20 1.325 1.725 2.086 2.528 2.845 3.850
21 1.323 1.721 2.080 2.518 2.831 3.819
22 1.321 1.717 2.074 2.508 2.819 3.792
23 1.319 1.714 2.069 2.500 2.807 3.767
24 1.318 1.711 2.064 2.492 2.797 3.745
25 1.316 1.708 2.060 2.485 2.787 3.725
26 1.315 1.706 2.056 2.479 2.779 3.707
27 1.314 1.703 2.052 2.473 2.771 3.690
28 1.313 1.701 2.048 2.467 2.763 3.674
29 1.311 1.699 2.045 2.462 2.756 3.659
30 1.310 1.697 2.042 2.457 2.750 3.646
40 1.303 1.684 2.021 2.423 2.704 3.551
60 1.296 1.671 2.000 2.390 2.660 3.460
120 1.289 1.658 1.980 2.358 2.617 3.373
1.282 1.645 1.960 2.326 2.576 3.291
N.B. When required df is not shown, use the next lowest number, except for very large dfs
(well over 120), when you can use the row for innity ().
Table D Critical values of
2
r
for three conditions at various levels of
probability (Friedman)
For any N the observed value of
2
r
is signicant at a given level of signicance if
it is equal to or larger than the critical values shown in Table D.
N = 2 N = 3 N = 4 N = 5
2
r
p
2
r
p
2
r
p
2
r
p
4 0.167 2.667 0.361 2.0 0.431 1.6 0.522
4.667 0.194 3.5 0.273 2.8 0.367
6.000 0.028 4.5 0.125 3.6 0.182
6.0 0.069 4.8 0.124
6.5 0.042 5.2 0.093
8.0 0.0046 6.4 0.039
7.6 0.024
8.4 0.0085
10.0 0.00077
N = 6 N = 7 N = 8 N = 9
2
r
p
2
r
p
2
r
p
2
r
p
4.00 0.184 3.429 0.237 3.00 0.285 2.667 0.328
4.33 0.142 3.714 0.192 3.25 0.236 2.889 0.278
5.33 0.072 4.571 0.112 4.00 0.149 3.556 0.187
6.33 0.052 5.429 0.085 4.75 0.120 4.222 0.154
7.00 0.029 6.000 0.052 5.25 0.079 4.667 0.107
8.33 0.012 7.143 0.027 6.25 0.047 5.556 0.069
9.00 0.0081 7.714 0.021 6.75 0.038 6.000 0.057
9.33 0.0055 8.000 0.016 7.00 0.030 6.222 0.048
10.33 0.0017 8.857 0.0084 7.75 0.018 6.889 0.031
12.00 0.00013 10.286 0.0036 9.00 0.0099 8.000 0.019
10.571 0.0027 9.25 0.0080 8.222 0.016
11.143 0.0012 9.75 0.0048 8.667 0.010
12.286 0.00032 10.75 0.0024 9.556 0.0060
14.000 0.000021 12.00 0.0011 10.667 0.0035
12.25 0.00086 10.889 0.0029
13.00 0.00026 11.556 0.0013
14.25 0.000061 12.667 0.00066
16.00 0.0000036 13.556 0.00035
14.000 0.00020
14.222 0.000097
14.889 0.000054
16.222 0.000011
18.000 0.0000006
Table E Critical values of
2
at various levels of probability (chi-square)
For any particular df the observed value of
2
is signicant at a given level of
df 0.10 0.05 0.02 0.01 0.001
1 2.71 3.84 5.41 6.64 10.83
2 4.60 5.99 7.82 9.21 13.82
3 6.25 7.82 9.84 11.34 16.27
4 7.78 9.49 11.67 13.28 18.46
5 9.24 11.07 13.39 15.09 20.52
6 10.64 12.59 15.03 16.81 22.46
7 12.02 14.07 16.62 18.48 24.32
8 13.36 15.51 18.17 20.09 26.12
9 14.68 16.92 19.68 21.67 27.88
10 15.99 18.31 21.16 23.21 29.59
11 17.28 19.68 22.62 24.72 31.26
12 18.55 21.03 24.05 26.22 32.91
13 19.81 22.36 25.47 27.69 34.53
14 21.06 23.68 26.87 29.14 36.12
15 22.31 25.00 28.26 30.58 37.70
16 23.54 26.30 29.63 32.00 39.29
17 24.77 27.59 31.00 33.41 40.75
18 25.99 28.87 32.35 34.80 42.31
19 27.20 30.14 33.69 36.19 43.82
20 28.41 31.41 35.02 37.57 45.32
21 29.62 32.67 36.34 38.93 46.80
22 30.81 33.92 37.66 40.29 48.27
23 32.01 35.17 38.97 41.64 49.73
24 33.20 36.42 40.27 42.98 51.18
25 34.38 37.65 41.57 44.31 52.62
26 35.56 38.88 42.86 45.64 54.05
27 36.74 40.11 44.14 46.97 55.48
28 37.92 41.34 45.42 48.28 56.89
29 39.09 42.56 46.69 49.59 58.30
30 40.26 43.77 47.96 50.89 59.70
Table F Critical values of H at various levels of probability
(KruskalWallis)
For any n
1
, n
2
, n
3
the observed value of H is signicant at a given level of
Size of groups Size of groups
n
1
n
2
n
3
H p n
1
n
2
n
3
H p
2 1 1 2.7000 0.500 4 3 1 5.8333 0.021
2 2 1 3.6000 0.200
5.2083 0.050
5.0000 0.057
2 2 2 4.5714 0.067 4.0556 0.093
3.7143 0.200 3.8889 0.129
3 1 1 3.2000 0.300 4 3 2 6.4444 0.008
3 2 1 4.2857 0.100
6.3000 0.011
3.8571 0.133
5.4444 0.046
5.4000 0.051
3 2 2 5.3572 0.029 4.5111 0.098
4.7143 0.048 4.4444 0.102
4.5000 0.067
4.4643 0.105 4 3 3 6.7455 0.010
6.7091 0.013
3 3 1 5.1429 0.043 5.7909 0.046
4.5714 0.100 5.7273 0.050
4.0000 0.129 4.7091 0.092
4.7000 0.101
3 3 2 6.2500 0.011
5.3611 0.032 4 4 1 6.6667 0.010
5.1389 0.061 6.1667 0.022
4.5556 0.100 4.9667 0.048
4.2500 0.121 4.8667 0.054
4.1667 0.082
3 3 3 7.2000 0.004 4.1667 0.082
6.4889 0.011
5.6889 0.029 4 4 2 7.0364 0.006
5.6000 0.050 6.8727 0.011
5.0667 0.086 5.4545 0.046
4.6222 0.100 5.2364 0.052
4 1 1 3.5714 0.200
4.5545 0.098
4.4455 0.103
4 2 1 4.8214 0.057
4.5000 0.076 4 4 3 7.1439 0.010
4.0179 0.114 7.1364 0.011
5.5985 0.049
4 2 2 6.0000 0.014 5.5758 0.051
5.3333 0.033 4.5455 0.099
5.1250 0.052 4.4773 0.102
4.4583 0.100
Table F (continued)
n
1
n
2
n
3
H p n
1
n
2
n
3
H p
4.1667 0.105 4 4 4 7.6538 0.008
7.5385 0.011
5.6923 0.049
5.6538 0.054
4.6539 0.097
4.5001 0.104
5 1 1 3.8571 0.143 5 4 3 7.4449 0.010
5 2 1 5.2500 0.036 7.3949 0.011
5.0000 0.048 5.6564 0.049
4.4500 0.071 5.6308 0.050
4.2000 0.095 4.5487 0.099
4.0500 0.119 4.5231 0.103
5 2 2 6.5333 0.008 5 4 4 7.7604 0.009
6.1333 0.013 7.7440 0.011
5.1600 0.034 5.6571 0.049
5.0400 0.056 5.6176 0.050
4.3733 0.090 4.6187 0.100
4.2933 0.122 4.5527 0.102
5 3 1 6.4000 0.012 5 5 1 7.3091 0.009
4.9600 0.048 6.8364 0.011
4.8711 0.052 5.1273 0.046
4.0178 0.095 4.9091 0.053
3.8400 0.123 4.1091 0.086
4.0364 0.105
5 3 2 6.9091 0.009 5 5 2 7.3385 0.010
6.8218 0.010 7.2692 0.010
5.2509 0.049 5.3385 0.047
5.1055 0.052 5.2462 0.051
4.6509 0.091 4.6231 0.097
4.4945 0.101 4.5077 0.100
5 3 3 7.0788 0.009 5 5 3 7.5780 0.010
6.9818 0.011 7.5429 0.010
5.6485 0.049 5.7055 0.046
5.5152 0.051 5.6264 0.051
4.5333 0.097 4.5451 0.100
4.4121 0.109 4.5363 0.102
5 4 1 6.9545 0.008 5 5 4 7.8229 0.010
6.8400 0.011 7.7914 0.010
4.9855 0.044 5.6657 0.049
4.8600 0.056 5.6429 0.050
3.9873 0.098 4.5229 0.099
3.9600 0.102 4.5200 0.101
n
1
n
2
n
3
H p n
1
n
2
n
3
H p
5 4 2 7.2045 0.009 5 5 5 8.0000 0.009
7.1182 0.010 7.9800 0.010
5.2727 0.049 5.7800 0.049
5.2682 0.050 5.6600 0.051
4.5409 0.098 4.5600 0.100
4.5182 0.101 4.5000 0.102
Table G Critical values of F at various levels of probability (ANOVA)
For any
1
and
2
the observed value of F is signicant at a given level of signicance if it is equal to or
larger than the critical values shown in Tables G(1)G(2).
Table G(1) Critical values of F at p < 0.05
2
1 2 3 4 5 6 7 8 10 12 24
5 6.61 5.79 5.41 5.19 5.05 4.95 4.88 4.82 4.74 4.68 4.53 4.36
6 5.99 5.14 4.76 4.53 4.39 4.28 4.21 4.15 4.06 4.00 3.84 3.67
7 5.59 4.74 4.35 4.12 3.97 3.87 3.79 3.73 3.64 3.57 3.41 3.23
8 5.32 4.46 4.07 3.84 3.69 3.58 3.50 3.44 3.35 3.28 3.12 2.93
9 5.12 4.26 3.86 3.63 3.48 3.37 3.29 3.23 3.14 3.07 2.90 2.71
10 4.96 4.10 3.71 3.48 3.33 3.22 3.14 3.07 2.98 2.91 2.74 2.54
11 4.84 3.98 3.59 3.36 3.20 3.09 3.01 2.95 2.85 2.79 2.61 2.40
12 4.75 3.89 3.49 3.26 3.11 3.00 2.91 2.85 2.75 2.69 2.51 2.30
13 4.67 3.81 3.41 3.18 3.03 2.92 2.83 2.77 2.67 2.60 2.42 2.21
14 4.60 3.74 3.34 3.11 2.96 2.85 2.76 2.70 2.60 2.53 2.35 2.13
15 4.54 3.68 3.29 3.06 2.90 2.79 2.71 2.64 2.54 2.48 2.29 2.07
16 4.49 3.63 3.24 3.01 2.85 2.74 2.66 2.59 2.49 2.42 2.24 2.01
17 4.45 3.59 3.20 2.96 2.81 2.70 2.61 2.55 2.45 2.38 2.19 1.96
18 4.41 3.55 3.16 2.93 2.77 2.66 2.58 2.51 2.41 2.34 2.15 1.92
19 4.38 3.52 3.13 2.90 2.74 2.63 2.54 2.48 2.38 2.31 2.11 1.88
20 4.35 3.49 3.10 2.87 2.71 2.60 2.51 2.45 2.35 2.28 2.08 1.84
21 4.32 3.47 3.07 2.84 2.68 2.57 2.49 2.42 2.32 2.25 2.05 1.81
22 4.30 3.44 3.05 2.82 2.66 2.55 2.46 2.40 2.30 2.23 2.03 1.78
23 4.28 3.42 3.03 2.80 2.64 2.53 2.44 2.37 2.27 2.20 2.00 1.76
24 4.26 3.40 3.01 2.78 2.62 2.51 2.42 2.36 2.25 2.18 1.98 1.73
25 4.24 3.39 2.99 2.76 2.60 2.49 2.40 2.34 2.24 2.16 1.96 1.71
26 4.23 3.37 2.98 2.74 2.59 2.47 2.39 2.32 2.22 2.15 1.95 1.69
27 4.21 3.35 2.96 2.73 2.57 2.46 2.37 2.31 2.20 2.13 1.93 1.67
28 4.20 3.34 2.95 2.71 2.56 2.45 2.36 2.29 2.19 2.12 1.91 1.65
29 4.18 3.33 2.93 2.70 2.55 2.43 2.35 2.28 2.18 2.10 1.90 1.64
30 4.17 3.32 2.92 2.69 2.53 2.42 2.33 2.27 2.16 2.09 1.89 1.62
32 4.15 3.29 2.90 2.67 2.51 2.40 2.31 2.24 2.14 2.07 1.86 1.59
34 4.13 3.28 2.88 2.65 2.49 2.38 2.29 2.23 2.12 2.05 1.84 1.57
36 4.11 3.26 2.87 2.63 2.48 2.36 2.28 2.21 2.11 2.03 1.82 1.55
38 4.10 3.24 2.85 2.62 2.46 2.35 2.26 2.19 2.09 2.02 1.81 1.53
40 4.08 3.23 2.84 2.61 2.45 2.34 2.25 2.18 2.08 2.00 1.79 1.51
60 4.00 3.15 2.76 2.53 2.37 2.25 2.17 2.10 1.99 1.92 1.70 1.39
120 3.92 3.07 2.68 2.45 2.29 2.18 2.09 2.02 1.91 1.83 1.61 1.25
3.84 3.00 2.60 2.37 2.21 2.10 2.01 1.94 1.83 1.75 1.52 1.00
N.B. When the required dfs are not shown use the next lowest number. For very large dfs (well over 120) you can use
innity ().
Table G(2) Critical values of F at p < 0.01
2
1 2 3 4 5 6 7 8 10 12 24
5 16.26 13.27 12.06 11.39 10.97 10.67 10.46 10.29 10.05 9.89 9.47 9.02
6 13.74 10.92 9.78 9.15 8.75 8.47 8.26 8.10 7.87 7.72 7.31 6.88
7 12.25 9.55 8.45 7.85 7.46 7.19 6.99 6.84 6.62 6.47 6.07 5.65
8 11.26 8.65 7.59 7.01 6.63 6.37 6.18 6.03 5.81 5.67 5.28 4.86
9 10.56 8.02 6.99 6.42 6.06 5.80 5.61 5.47 5.26 5.11 4.73 4.31
10 10.04 7.56 6.55 5.99 5.64 5.39 5.20 5.06 4.85 4.71 4.33 3.91
11 9.65 7.21 6.22 5.67 5.32 5.07 4.89 4.74 4.54 4.40 4.02 3.60
12 9.33 6.93 5.95 5.41 5.06 4.82 4.64 4.50 4.30 4.16 3.78 3.36
13 9.07 6.70 5.74 5.21 4.86 4.62 4.44 4.30 4.10 3.96 3.59 3.17
14 8.86 6.51 5.56 5.04 4.70 4.46 4.28 4.14 3.94 3.80 3.43 3.00
15 8.68 6.36 5.42 4.89 4.56 4.32 4.14 4.00 3.80 3.67 3.29 2.87
16 8.53 6.23 5.29 4.77 4.44 4.20 4.03 3.89 3.69 3.55 3.18 2.75
17 8.40 6.11 5.18 4.67 4.34 4.10 3.93 3.79 3.59 3.46 3.08 2.65
18 8.29 6.01 5.09 4.58 4.25 4.01 3.84 3.71 3.51 3.37 3.00 2.57
19 8.18 5.93 5.01 4.50 4.17 3.94 3.77 3.63 3.43 3.30 2.92 2.49
20 8.10 5.85 4.94 4.43 4.10 3.87 3.70 3.56 3.37 3.23 2.86 2.42
21 8.02 5.78 4.87 4.37 4.04 3.81 3.64 3.51 3.31 3.17 2.80 2.36
22 7.95 5.72 4.82 4.31 3.99 3.76 3.59 3.45 3.26 3.12 2.75 2.31
23 7.88 5.66 4.76 4.26 3.94 3.71 3.54 3.41 3.21 3.07 2.70 2.26
24 7.82 5.61 4.72 4.22 3.90 3.67 3.50 3.36 3.17 3.03 2.66 2.21
25 7.77 5.57 4.68 4.18 3.86 3.63 3.46 3.32 3.13 2.99 2.62 2.17
26 7.72 5.53 4.64 4.14 3.82 3.59 3.42 3.29 3.09 2.96 2.58 2.13
27 7.68 5.49 4.60 4.11 3.78 3.56 3.39 3.26 3.06 2.93 2.55 2.10
28 7.64 5.45 4.57 4.07 3.75 3.53 3.36 3.23 3.03 2.90 2.52 2.06
29 7.60 5.42 4.54 4.04 3.73 3.50 3.33 3.20 3.00 2.87 2.49 2.03
30 7.56 5.39 4.51 4.02 3.70 3.47 3.30 3.17 2.98 2.84 2.47 2.01
32 7.50 5.34 4.46 3.97 3.65 3.43 3.26 3.13 2.93 2.80 2.42 1.96
34 7.45 5.29 4.42 3.93 3.61 3.39 3.22 3.09 2.90 2.76 2.38 1.91
36 7.40 5.25 4.38 3.89 3.58 3.35 3.18 3.05 2.86 2.72 2.35 1.87
38 7.35 5.21 4.34 3.86 3.54 3.32 3.15 3.02 2.83 2.69 2.32 1.84
40 7.31 5.18 4.31 3.83 3.51 3.29 3.12 2.99 2.80 2.66 2.29 1.80
60 7.08 4.98 4.13 3.65 3.34 3.12 2.95 2.82 2.63 2.50 2.12 1.60
120 6.85 4.79 3.95 3.48 3.17 2.96 2.79 2.66 2.47 2.34 1.95 1.38
6.63 4.61 3.78 3.32 3.02 2.80 2.64 2.51 2.32 2.18 1.79 1.00
Table H Critical values of r at various levels of probability (Pearson
correlation)
The observed value of r is signicant at a given level of signicance if it is equal to
or larger than the critical values shown in the table
Level of signicance for one-tailed test
.05 .025 .01 .005 .0005
Level of signicance for two-tailed test
df = N 2 .10 .05 .02 .01 .001
1 .9877 .9969 .9995 .9999 1.0000
2 .9000 .9500 .9800 .9900 .9990
3 .8054 .8783 .9343 .9587 .9912
4 .7293 .8114 .8822 .9172 .9741
5 .6694 .7545 .8329 .8745 .9507
6 .6215 .7067 .7887 .8343 .9249
7 .5822 .6664 .7498 .7977 .8982
8 .5494 .6319 .7155 .7646 .8721
9 .5214 .6021 .6851 .7348 .8471
10 .4973 .5760 .6581 .7079 .8233
11 .4762 .5529 .6339 .6835 .8010
12 .4575 .5324 .6120 .6614 .7800
13 .4409 .5139 .5923 .6411 .7603
14 .4259 .4973 .5742 .6226 .7420
15 .4124 .4821 .5577 .6055 .7246
16 .4000 .4683 .5425 .5897 .7084
17 .3887 .4555 .5285 .5751 .6932
18 .3783 .4438 .5155 .5614 .6787
19 .3687 .4329 .5034 .5487 .6652
20 .3598 .4227 .4921 .5368 .6524
25 .3233 .3809 .4451 .4869 .5974
30 .2960 .3494 .4093 .4487 .5541
35 .2746 .3246 .3810 .4182 .5189
40 .2573 .3044 .3578 .3932 .4896
45 .2428 .2875 .3384 .3721 .4648
50 .2306 .2732 .3218 .3541 .4433
60 .2108 .2500 .2948 .3248 .4078
70 .1954 .2319 .2737 .3017 .3799
80 .1829 .2172 .2565 .2830 .3568
90 .1726 .2050 .2422 .2673 .3375
100 .1638 .1946 .2301 .2540 .3211
N.B. When there is no exact df use the next lowest number.

Learning To Use Statistical Tests in Psychology

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Learning To Use Statistical Tests in Psychology

Uploaded by

Copyright:

Available Formats

Learning to use

Progress box one

Progress box two

Progress box three

Progress box four

Progress box six

Progress box seven

Progress box eight

d) is calculated taking into account the plus

d) represents the predicted differences between the conditions

Progress box nine

d), taking into account plus and

(450 409.6) + (165 136.9)

Progress box ten

(121 + 441 + 256)

Progress box eleven

Progress box twelve

Progress box thirteen

Progress box fourteen

Progress box fteen

Progress box seventeen

Progress box eighteen

Progress box nineteen

Progress box twenty

b = totals for each variable

Progress box twenty one

Progress box twenty two

Progress box twenty three

You might also like